<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Furkan Gözükara - PhD Computer Engineer, SECourses on Medium]]></title>
        <description><![CDATA[Stories by Furkan Gözükara - PhD Computer Engineer, SECourses on Medium]]></description>
        <link>https://medium.com/@furkangozukara?source=rss-fb0739c3d821------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*2Ccv7W6iWnRFKicvDFVY_Q.jpeg</url>
            <title>Stories by Furkan Gözükara - PhD Computer Engineer, SECourses on Medium</title>
            <link>https://medium.com/@furkangozukara?source=rss-fb0739c3d821------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 17:10:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@furkangozukara/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[New MiniMax H3 Image Preset improves errors in tiny details significantly]]></title>
            <link>https://medium.com/@furkangozukara/new-minimax-h3-image-preset-improves-errors-in-tiny-details-significantly-54c1e2b0464b?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/54c1e2b0464b</guid>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 03:43:49 GMT</pubDate>
            <atom:updated>2026-11-15T03:43:49.131Z</atom:updated>
            <content:encoded><![CDATA[<h3>Zoom In at 16 MP: A Real Ratchet Instead of a Fake Tool With the New MiniMax H3 Image Preset</h3><h3>Full tutorial link &gt; <a href="https://www.youtube.com/watch?v=0EkDXouIiDA">https://www.youtube.com/watch?v=0EkDXouIiDA</a></h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F0EkDXouIiDA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D0EkDXouIiDA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F0EkDXouIiDA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9d171e14c9c9fd2aabe51c325ec1ec1a/href">https://medium.com/media/9d171e14c9c9fd2aabe51c325ec1ec1a/href</a></iframe><h3>Info</h3><p>MiniMax H3 Image makes 16 megapixel pictures in only 6 steps on your own NVIDIA GPU, and it now has a second text to image preset for small objects: “MiniMax H3 Image — Text To Image 16 MP — KJNodes Empty Image — 6 Steps — 261114”. It ships ready beside the standard preset in our SECourses ComfyUI installer V152 and in Amazing SwarmUI Presets v90, and in SwarmUI it is one checkbox of our MiniMaxH3Image extension. On many images it draws small objects more correctly: gem settings, watch gears, tools, bike sprockets and plated dishes. The model, 16 MP, 6 steps and MiniMax’s official prompt enhancer running locally stay the same. Faces and rooms often look glossier and more plastic with it, so the standard preset stays the choice for portraits and natural skin.</p><h3>Links:</h3><ul><li>Download the ComfyUI installer zip V152 with the new KJNodes Empty Image preset (our ComfyUI post): [ <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a> ]</li><li>SwarmUI Model Downloader v193 with Amazing SwarmUI Presets v90 and the new preset (our SwarmUI post): [ <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a> ]</li><li>Full ComfyUI installation tutorial: [ <a href="https://youtu.be/nxsCBTcCVp4">https://youtu.be/nxsCBTcCVp4</a> ]</li><li>SwarmUI full tutorial: [ <a href="https://youtu.be/0HmIeeI1iBo">https://youtu.be/0HmIeeI1iBo</a> ]</li><li>How to update ComfyUI, SwarmUI, models and presets: [ <a href="https://www.youtube.com/watch?v=sdD3--WbHZk">https://www.youtube.com/watch?v=sdD3--WbHZk</a> ]</li><li>One-time Windows requirements installer tutorial: [ <a href="https://youtu.be/ptuLL0C72Ms">https://youtu.be/ptuLL0C72Ms</a> ]</li><li>SECourses Discord: [ <a href="https://discord.com/invite/software-engineering-courses-secourses-772774097734074388">https://discord.com/invite/software-engineering-courses-secourses-772774097734074388</a> ]</li></ul><p>You will learn which preset fits which picture, from same-caption, same-seed pairs: each comparison shows the whole 16 MP picture first and then zooms into the small part where the two presets differ. In ComfyUI you open the new preset from the Presets folder with Ctrl+O, see that the prompt box, output size, prompt enhancer and seed number match the standard preset, and find the one difference: the MiniMax H3 Prompt Weight node of KJNodes in “text with empty image” mode, which shows the model an empty black picture beside your text for the first quarter of the steps (2 of 6); the other 4 steps run on your text alone. Then you type a short idea, keep the enhancer on, click Run and look closely at the watch movement of the 16 MP result. In SwarmUI you filter the Presets tab for H3 Image, reset the parameters in Quick Tools, Direct Apply the new card and find its KJNodes Empty Image checkbox, which also works with your own settings. Finally you generate a jeweler at 16 MP and zoom into the ring.</p><p>Main topics include MiniMax H3 Image, KJNodes, ComfyUI and SwarmUI presets, text to image, small object detail, jewelry, watch and food images, the local prompt enhancer and 16 MP.</p><p>Use the chapters to jump to the comparisons, the ComfyUI steps or the SwarmUI steps.</p><h3>Chapters:</h3><ul><li><a href="https://youtu.be/0EkDXouIiDA?t=0">00:00:00</a> New preset fixes a jammed diamond setting at 16 MP</li><li><a href="https://youtu.be/0EkDXouIiDA?t=24">00:00:24</a> Small objects compared: watch gears, ratchet, sprockets, dishes</li><li><a href="https://youtu.be/0EkDXouIiDA?t=55">00:00:55</a> The trade-off: glossier, more plastic faces and rooms</li><li><a href="https://youtu.be/0EkDXouIiDA?t=66">00:01:06</a> Which preset to use for which image</li><li><a href="https://youtu.be/0EkDXouIiDA?t=91">00:01:31</a> ComfyUI: open the KJNodes Empty Image preset</li><li><a href="https://youtu.be/0EkDXouIiDA?t=111">00:01:51</a> The one new box: KJNodes empty black picture</li><li><a href="https://youtu.be/0EkDXouIiDA?t=127">00:02:07</a> Generate a watchmaker image with the new preset</li><li><a href="https://youtu.be/0EkDXouIiDA?t=172">00:02:52</a> Inspect the watch gears in the 16 MP result</li><li><a href="https://youtu.be/0EkDXouIiDA?t=185">00:03:05</a> SwarmUI: apply the new preset card</li><li><a href="https://youtu.be/0EkDXouIiDA?t=210">00:03:30</a> The KJNodes Empty Image switch in SwarmUI</li><li><a href="https://youtu.be/0EkDXouIiDA?t=220">00:03:40</a> Generate a jeweler image and zoom into the diamond</li><li><a href="https://youtu.be/0EkDXouIiDA?t=268">00:04:28</a> Recap: which preset for which image</li></ul><h3>Extra Info</h3><p>Method credit: the “text with empty image” mode of the MiniMax H3 Prompt Weight node in KJNodes by Kijai. We tested it against our standard preset on 40 same-caption, same-seed pairs at 16 MP and ship it as a second preset.</p><p>Music: Champion by Telecasted, Ghibli Station by The Mini Vandals, Curadora by Casa Rosa, Chat Roulette by Single Friend, Swing Haven 4 — Tulsa by Reed Mathis, Vespers by Topher Mohr and Alex Elena, On The Beach by Telecasted, TrueArtRealAffectionPart4 by Noir Et Blanc Vie.</p><p>Which preset when: the new preset when small objects carry the picture (watch movements, rings and gem settings, tools, sprockets, plated dishes); the standard preset for portraits, natural skin and interiors that should look like a photograph. For an important image, run both with the same prompt and keep the better one. The new preset is text to image only; for photo edits keep the Image Edit preset. No model yet? The MiniMax-H3 Core Bundle in the Model Downloader fetches every file; the Low VRAM Bundle needs less GPU memory.</p><p>This video is for ComfyUI and SwarmUI users who make product, jewelry, food and mechanism images with MiniMax H3 Image. New to ComfyUI or SwarmUI? Start with the full tutorials linked above; already installed, the update tutorial shows how to get the new presets.</p><p>For support, questions, feature requests or installation issues, see the pinned comment, our Patreon posts, our Discord and the video comments. Thank you for watching.</p><h3>14 November 2026 Update V152 for ComfyUI</h3><p>New: MiniMax H3 Image with KJNodes’ empty-image method, for cleaner skin, hands and fine detail. A new Text To Image preset shows the model an empty picture during the first steps, a method from kijai’s KJNodes. On many prompts it gives more natural skin, cleaner hands and sharper small objects than the standard preset, at the same 16 MP and with the same prompt enhancer.</p><p>The prompt enhancer is now repeatable, too: a fixed seed gives the same caption and the same image every time, in every MiniMax H3 Image preset.</p><p>The complete Presets folder of V152, the new preset marked NEW and the improved ones UPDATED:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*U_EhvTXvU-aNOvYm.png" /></figure><p>Every image in this post comes from a real run of the shipped presets with their default settings.</p><h3>NEW PRESET — TEXT TO IMAGE WITH KJNODES EMPTY IMAGE</h3><p>“MiniMax H3 Image — Text To Image 16 MP — KJNodes Empty Image — 6 Steps — 261114”</p><p>Cleaner skin, hands and fine detail from the same short prompt. Write your idea and press Run, exactly like the standard Text To Image preset. It uses the same models, all in the MiniMax-H3 Core Bundle:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RGcT2XAj1FTYverp.png" /></figure><p>Already using MiniMax H3 Image? There is nothing new to download. The MiniMax-H3 Low VRAM Bundle has the smaller W4A8 versions.</p><p>The same caption and seed, the standard preset on the left and the new one on the right, with details at their real pixels below. A watchmaker at his bench: the watch case, the gears and the loupe come out clean:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*q1GMsPwEvGOoDNW2.jpeg" /></figure><p>A chef plating a dish for a restaurant ad, with clean hands and a detailed plate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*D4FhM_bJkczhRiim.jpeg" /></figure><p>An olive farmer for a brand campaign: natural, weathered skin instead of a crackled texture:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sve1pdUs1mhRlGLF.jpeg" /></figure><p>The empty picture also changes the composition, like a new seed, so keep both presets and use the one you like for each prompt.</p><p>Where to find it: open the preset, write your idea in the gallery box and press Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*l64hRIPVUeGmf0zw.png" /></figure><p>The box under the result shows the exact prompt the model received. This preset makes new images; to edit your photos, use the Image Edit preset.</p><h3>UPDATED — MINIMAX H3 IMAGE TEXT TO IMAGE AND IMAGE EDIT</h3><ul><li>A fixed seed repeats your image: the prompt enhancer now writes the same caption for the same seed, so the image repeats too.</li><li>Tidier layout: the model loaders no longer overlap the nodes next to them.</li></ul><h3>UPDATED — KREA 2, PHANTOM WAN, LTX AND FLUX PRESETS</h3><ul><li>Krea 2: all three presets use the BF16 text encoder of the Krea 2 Core Bundle under its own file name. Already using Krea 2? Download the Krea 2 Core Bundle once more with SwarmUI Model Downloader V193 to add it. Raw Base also has a negative prompt box now.</li><li>Phantom Wan 14B Character Reference and LTX2.3 Licon MSR V2: the reference image switches open with the right on and off settings.</li><li>FLUX 2 Low RAM: the VRAM cleanup before sampling is optional now and off by default; switch it on if you run out of VRAM.</li><li>LTX2.5 Text or Image To Video and FLUX 3 Klein Advanced Image Edit: they open without missing example images.</li></ul><h3>ONE INSTALLER FOR EVERY PRESET</h3><ul><li>Windows_Install_Or_Update_ComfyUI.bat installs every custom node our presets use, and so do the RunPod and Massed Compute installers: no bundle menu, no separate node installer.</li><li>IP-Adapter and FaceID models now come from the new IP-Adapter And FaceID Bundle in our SwarmUI Model Downloader.</li><li>Installer fixes: a moved, renamed or copied ComfyUI folder gets a working environment again, and a failed step is reported as an error.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V152.zip from this post and extract it over your installer folder.</li><li>Close ComfyUI, and SwarmUI too if it uses this ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat. It installs and updates every node our presets use.</li><li>If you do not have it yet, download the MiniMax-H3 Core Bundle (or the Low VRAM Bundle) with our SwarmUI Model Downloader V190 or newer.</li><li>Restart ComfyUI and refresh the browser tab.</li></ol><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>14 November 2026 Update V193 for SwarmUI</h3><p>New in SwarmUI: MiniMax H3 Image with KJNodes’ empty-image method, for cleaner skin, hands and fine detail. A new preset shows the model an empty picture during the first steps, a method from kijai’s KJNodes. On many prompts it gives more natural skin, cleaner hands and sharper small objects than the standard Text To Image preset, at the same 16 MP and with the same prompt enhancer. It is also a new switch, MiniMax H3 Image KJNodes Empty Image, so you can turn it on with your own settings.</p><p>The prompt enhancer is now repeatable, too: a fixed seed gives the same caption and the same image every time.</p><p>Amazing SwarmUI Presets moves to V90 with 102 presets: 1 new and 5 updated.</p><p>All 102 presets in the Presets tab, the new one marked NEW and the updated ones UPDATED:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6BgCL8aMpvgINqBW.jpeg" /></figure><p>The new and updated presets up close; the new one has its own great hornbill thumbnail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*293fQnPei517SfIc.png" /></figure><p>Windows_Preset_Delete_Import.bat brings all of them into your SwarmUI; the steps are in How to update at the end.</p><h3>NEW PRESET — TEXT TO IMAGE WITH KJNODES EMPTY IMAGE</h3><p><em>“MiniMax H3 Image — Text To Image 16 MP — KJNodes Empty Image — 6 Steps — 261114”</em></p><p>Cleaner skin, hands and fine detail from the same short prompt. Apply it, write your idea and click Generate, exactly like the standard Text To Image preset. It uses the same models, all in the MiniMax-H3 Core Bundle:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CVNcrGNRF1gEVaXn.png" /></figure><p>Already using MiniMax H3 Image? There is nothing new to download. The MiniMax-H3 Low VRAM Bundle has the smaller W4A8 versions.</p><p>The same caption and seed, the standard preset on the left and the new one on the right, made with the same presets in our ComfyUI, with details at their real pixels below. A watchmaker at his bench: the watch case, the gears and the loupe come out clean:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hRKqIKw3EMgCmevb.jpeg" /></figure><p>A chef plating a dish for a restaurant ad, with clean hands and a detailed plate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*G3s_inp7_w4xgzH_.jpeg" /></figure><p>An olive farmer for a brand campaign: natural, weathered skin instead of a crackled texture:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rNlDpOvBTHnApU9_.jpeg" /></figure><p>The empty picture also changes the composition, like a new seed, so keep both presets and use the one you like for each prompt.</p><p>Where to find it: Presets tab → filter 261114 → apply the preset, write your idea and click Generate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nT73wPeU4r62lH8k.png" /></figure><p>The switch works for text to image; when you attach images to the prompt, the preset edits them as usual.</p><h3>UPDATED — MINIMAX H3 IMAGE TEXT TO IMAGE AND IMAGE EDIT</h3><ul><li>A fixed seed repeats your image: the prompt enhancer now writes the same caption for the same seed, so the image repeats too.</li><li>Both presets keep the new switch off, so they work as before.</li></ul><h3>UPDATED — KREA 2</h3><p>All three Krea 2 presets now select the BF16 text encoder of the Krea 2 Core Bundle under its own file name, so they always get it instead of SwarmUI’s smaller default. Already using Krea 2? Download the Krea 2 Core Bundle once more to add it.</p><h3>NEW IN THE MODEL DOWNLOADER</h3><ul><li>IP-Adapter And FaceID Bundle (SD 1.5, SDXL, Kolors): every IP-Adapter, FaceID adapter, FaceID LoRA and image encoder in one click, for SwarmUI’s IP-Adapter feature and for ComfyUI.</li><li>Wan 2.2 Core 4 Steps Bundle now includes RIFE 4.7 frame interpolation.</li><li>LTX 2.5 Upscaler bundle: the smaller INT8 text encoder that SwarmUI uses by default, plus the Gemma 4 E2B prompt enhancer.</li><li>Download queue: shows how many downloads have finished, failed or been cancelled.</li><li>SwarmUI installer: if SwarmUI cannot be installed, it says why and what to do next.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v193.zip from this post and extract it over your downloader folder.</li><li>Close SwarmUI and run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI and our extensions.</li><li>Update the ComfyUI backend: close SwarmUI again (step 2 starts it) and run Windows_Install_Or_Update_ComfyUI.bat from our ComfyUI V152.</li><li>Restart SwarmUI, then run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v90.json. It backs up your current presets first, then replaces them all, your own presets too.</li><li>If you do not have it yet, run Windows_Start_Download_Models_App.bat and download the MiniMax-H3 Core Bundle, or the MiniMax-H3 Low VRAM Bundle for less memory.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=54c1e2b0464b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MiniMax H3 Image Is Here: 16 MP Images and Edits in 6 Steps, Running Locally in ComfyUI — Ultra…]]></title>
            <link>https://medium.com/@furkangozukara/minimax-h3-image-is-here-16-mp-images-and-edits-in-6-steps-running-locally-in-comfyui-ultra-978de93669a2?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/978de93669a2</guid>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 02:51:17 GMT</pubDate>
            <atom:updated>2026-11-11T02:51:17.590Z</atom:updated>
            <content:encoded><![CDATA[<h3>MiniMax H3 Image Is Here: 16 MP Images and Edits in 6 Steps, Running Locally in ComfyUI — Ultra fast even at 16 MP and works as low as 6 GB VRAM</h3><h3>1 MP vs 16 MP: See Every Whisker With MiniMax H3 Image in ComfyUI, Text to Image in 6 Steps</h3><h3>Full tutorial link &gt; <a href="https://www.youtube.com/watch?v=xCtDQOn2zWQ">https://www.youtube.com/watch?v=xCtDQOn2zWQ</a></h3><ul><li>Download the ComfyUI installer zip V150 with both MiniMax H3 Image presets (our ComfyUI post): [ <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a> ]</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FxCtDQOn2zWQ%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DxCtDQOn2zWQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FxCtDQOn2zWQ%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/67a14fb9a1a1d98c761cbc5c509efd2c/href">https://medium.com/media/67a14fb9a1a1d98c761cbc5c509efd2c/href</a></iframe><p>MiniMax H3 Image is a new 33.1B image model that generates and edits pictures at 16 megapixels, 4096 × 4096 or a matching wide or tall size, in only 6 sampling steps on your own NVIDIA GPU, with no upscaler. Our SECourses ComfyUI installer V150 ships two ready presets for it, Text To Image 16 MP and Image Edit Up To 16 Images, with every model and setting already chosen. MiniMax’s official prompt enhancer runs locally inside them in three modes (reasoning, fast and off): it turns a short idea in any language into the detailed caption the model was trained on, picks the aspect ratio, and a box shows the exact prompt the model received. The SECourses Model Downloader V190 fetches every model file into the right folders from one bundle. In this mini tutorial a rusted, moss-covered Volkswagen Beetle is restored with one plain sentence, 9 typed words become a 16 MP photo of a fox, and you see a room furnished from 4 photos, letter-perfect posters and 1 MP vs 4 MP vs 16 MP.</p><p>Links:</p><ul><li>Download the ComfyUI installer zip V150 with both MiniMax H3 Image presets (our ComfyUI post): [ <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a> ]</li><li>SECourses Model Downloader V190 with the MiniMax-H3 Core and Low VRAM bundles (our SwarmUI post): [ <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a> ]</li><li>Full ComfyUI installation tutorial: [ <a href="https://youtu.be/nxsCBTcCVp4">https://youtu.be/nxsCBTcCVp4</a> ]</li><li>How to update ComfyUI (mini tutorial): [ <a href="https://www.youtube.com/watch?v=sdD3--WbHZk">https://www.youtube.com/watch?v=sdD3--WbHZk</a> ]</li><li>ComfyUI mini tutorials playlist: [ <a href="https://www.youtube.com/playlist?list=PLABQ076bsdXU">https://www.youtube.com/playlist?list=PLABQ076bsdXU</a> ]</li><li>SECourses Discord: [ <a href="https://discord.com/invite/software-engineering-courses-secourses-772774097734074388">https://discord.com/invite/software-engineering-courses-secourses-772774097734074388</a> ]</li></ul><p>You will learn how to open the MiniMax H3 Image text to image preset from the Presets folder with Ctrl+O, write a short idea in box 1 in any language, keep the recommended 16 MP output size, use the prompt enhancer in reasoning mode for the best captions or fast mode for quick tries, switch it off when you write a long caption yourself, read the exact prompt the model received, and inspect the result at full size. In the image edit preset you add your photo to the gallery as @image1, write the change as one plain sentence with no mask, say what must stay the same (the forest, the light, the camera angle), and name more reference images with @image2 and @image3. Finally you set up the route yourself: download the ComfyUI V150 zip from our Patreon post, extract it over your ComfyUI folder, run the update file, install bundle 100 with the bundles installer, then open the MiniMax-H3 Core Bundle in the Model Downloader (4 files for images, about 44 GB) or the Low VRAM Bundle, which brings a smaller version of the image model.</p><p>Main topics include MiniMax H3 Image, ComfyUI presets, text to image, AI photo restoration, multi-image editing with @image tags, the local prompt enhancer, automatic aspect ratio, 6-step sampling, native 16 MP without an upscaler, RTX 5090 speed and lower VRAM use.</p><p>Use the chapters to jump to the first image, the photo edit, the comparisons, the speed numbers or the downloads.</p><h3>Chapters:</h3><ul><li><a href="https://youtu.be/xCtDQOn2zWQ?t=0">00:00:00</a> A rusted wreck restored at 16 megapixels</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=81">00:01:21</a> Where the two new MiniMax presets live</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=97">00:01:37</a> Text to image: a 16 MP picture from words</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=142">00:02:22</a> The prompt enhancer: MiniMax’s own, running locally</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=202">00:03:22</a> The exact prompt the model received</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=229">00:03:49</a> Fast enhancer mode: a 16 MP image in real time</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=277">00:04:37</a> Image edit: restore a photo with one sentence</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=334">00:05:34</a> The restored car and why it moved into a showroom</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=371">00:06:11</a> Many references, lettering, sizes, speed and lower VRAM</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=402">00:06:42</a> 1 vs 4 vs 16 MP: 16 MP recommended</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=418">00:06:58</a> RTX 5090 speed, and with a 12 GB limit</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=447">00:07:27</a> Get the V150 zip and the model bundle</li><li><a href="https://youtu.be/xCtDQOn2zWQ?t=479">00:07:59</a> The MiniMax-H3 model bundle in the Model Downloader</li></ul><p>Measured numbers and their conditions: on an RTX 5090 with the models loaded, 6 steps, prompt encoding and decoding included and the prompt enhancer not included, 1 MP took 3.4 s, 4 MP 5.7 s and 16 MP 15.7 s. With ComfyUI limited to 12 GB of the same card’s memory, a 16 MP image still finished in 19.5 s, the rest streaming from system RAM; a real 12 GB card computes slower. On camera, the fast enhancer mode made a second 16 MP fox in under half a minute from the click, and reasoning mode adds about a minute for the best captions.</p><p>This video is for ComfyUI users and AI image makers with an NVIDIA GPU who want the newest local image model at its full 16 megapixel quality without collecting files, nodes and settings by hand: open a preset, write one sentence and click Run. If ComfyUI is new for you, start with the installation tutorial linked above; if it is already installed, the update tutorial shows the zip update step by step.</p><p>For support, questions, feature requests or installation issues, see the pinned comment, our Patreon posts, our Discord and the video comments. Thank you for watching.</p><h3>10 November 2026 Update V150</h3><p>Presets Folder Content In Latest Zip File</p><p>MiniMax H3 Image is here: 16-megapixel images in 6 steps. Two new presets turn a few words into images up to 6272 pixels wide and edit your photos with up to 16 images at once. MiniMax’s own prompt enhancer runs locally inside the presets: a short idea in any language becomes the detailed caption the model was trained on, and the enhancer even picks the best shape for your image.</p><p>Real runs of the new presets:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*I_6M-xE3dt5hxEXf.jpeg" /></figure><p>Every image in this post comes from a real run of the shipped presets with their default settings.</p><p>The complete Presets folder of V150, the new ones marked NEW:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9Ia1D6bMqgaDQm3h.png" /></figure><p>To run them, download the MiniMax-H3 Core Bundle. It now includes MiniMax H3 Image, its VAE and the prompt enhancer model; the text encoder is the one MiniMax H3 video already uses, so H3 users download only the new files:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FkhJxUcR3UvJ_BTt.png" /></figure><p>Both new presets use this one bundle. The MiniMax-H3 Low VRAM Bundle has the smaller W4A8 versions.</p><h3>NEW PRESET — TEXT TO IMAGE AT 16 MEGAPIXELS</h3><p>“MiniMax H3 Image — Text To Image 16 MP — 6 Steps — 261110”</p><p>Type an idea, press Run, get a 16-megapixel image. The prompt enhancer writes the full description and chooses the aspect ratio, from tall posters to 21:9 panoramas. A snow leopard from one sentence, at 5024 x 3360:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bkSdIqV8orNEwFxf.jpeg" /></figure><p>It writes text you can read. A travel poster with two exact lines of lettering, 3072 x 5472:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aNdYGUwLCDBoAz7o.jpeg" /></figure><p>Panoramas come out in one piece, with no upscaler. Asked for a wide view, the enhancer chose 21:9, 6272 x 2688; below it, three details at their real pixels:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*h_f2_GSczbRTpaRd.jpeg" /></figure><p>Fine detail at 4096 x 4096. Open it full size and read the screw slots and the jewels:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*a-UjiI89v7ydhiNh.jpeg" /></figure><p>Where to find it: open the preset, write your idea in the gallery box and press Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CGKNauv9EoWa9FIQ.png" /></figure><p>The box under the result shows the exact prompt the model received. Set the enhancer to Off to use your own long caption; our new MiniMax H3 Image LLM guide in the zip makes ChatGPT, Claude or Gemini write one for you.</p><h3>1 MP VS 4 MP VS 16 MP</h3><p>Why 16 MP is the default. The same caption and seed at three sizes, with the face from each cropped to the same share of the frame:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UfSbBjp7C72CMw6U.jpeg" /></figure><p>16 MP has sixteen times the pixels of 1 MP, and it uses them: fur, whiskers and frost that 1 MP can only blur. It stays fast, too. Seconds per image on an RTX 5090, also with ComfyUI limited to 24, 16 or 12 GB of its memory:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FiC6-jaZk-ZWPPwi.png" /></figure><p>A 16 MP image takes about 16 seconds with all memory and about 20 seconds when ComfyUI may use only 12 GB; the rest streams from system RAM. Lower the size in 02 OUTPUT SIZE whenever you want a quick draft.</p><h3>NEW PRESET — IMAGE EDIT WITH UP TO 16 IMAGES</h3><p>“MiniMax H3 Image — Image Edit Up To 16 Images — 6 Steps — 261110”</p><p>Edit your photo with one sentence. Add the image, write the change in any language and press Run. The prompt enhancer looks at your images and writes the full edit description, and the result keeps your photo’s shape at 16 MP. A moss-covered wreck comes back to life, while the forest, the branches and the camera stay:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vcC97ahX446ey7KI.jpeg" /></figure><p>Combine many images. An empty room and three product photos, named @image1 to @image4 in the instruction: the exact armchair, lamp and plant land in the unchanged room:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kQdQwNGEFbJcTr-e.jpeg" /></figure><p>Where to find it: open the preset, add your images to the gallery, write the change and press Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zAe65WJR9ygcOLXG.png" /></figure><p>Leave the gallery empty and it works as a text to image preset, so both presets do both jobs.</p><h3>BUILT IN, LIKE ALL OUR PRESETS</h3><ul><li>Batches: + Prompt adds more prompts in a row, and a folder of .txt files runs one image per file.</li><li>The prompt enhancer in three modes: official with reasoning (best), fast without reasoning, or off.</li><li>Optional LoRA stack, shared seed and images per prompt.</li><li>Less memory: pick the W4A8 model in the same loader; BF16 is there for original precision.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V150.zip from this post and extract it over your installer folder.</li><li>Close ComfyUI, and SwarmUI too if it uses this ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install bundle 100.</li><li>Download the MiniMax-H3 Core Bundle (or the Low VRAM Bundle) with our SwarmUI Model Downloader V190 or newer.</li><li>Restart ComfyUI and refresh the browser tab.</li></ol><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>4 November 2026 Update V149</h3><p>Presets Folder Content In Latest Zip File</p><p>Edit real videos with MiniMax H3, and let the presets do the work. Six new presets replace the background of any video from one image, relight a whole clip, put a new face on someone, move a subject along a path you draw, upscale your video and replace a person or object while every other pixel stays as filmed. The Pika H3 Flex presets now write their own prompt when you leave it empty, and a long video can follow a story, scene by scene.</p><p>Real footage and real runs of the shipped presets:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FUtH2QhekKjy-_p3.jpeg" /></figure><p>Every result in this post comes from a real run of the shipped presets.</p><p>The complete Presets folder of V149, the new ones marked NEW and the updated ones UPDATED:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qg0K-Gf4xV6eYLuT.png" /></figure><p>To run them, download the MiniMax-H3 Core Bundle. It includes the FaceSwap LoRA and the H3 latent upscaler; if you already have the bundle, only the missing files download:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dSVK1kbS8Nj8zbgY.png" /></figure><p>All the new presets and the updated Pika H3 Flex presets use this one bundle.</p><h3>BUILT IN — OUR LIVE TOKEN COUNTER</h3><p>See what a run needs before you press Run. Our gallery counts the tokens of your text, references, audio and video as you work and shows them against the MiniMax H3 limit, with the size, the length and the mode beside them. It moves from green to amber to red as you get close, so you can shorten a clip or lower the size first. With long video windows it counts per window, because each window is one run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xBZud8m_Si4_xVxt.png" /></figure><p>Beside it in every preset: Load + trim to pick the part of your clip, folder batch with optional merging, and Continue From Last Video Frames.</p><h3>NEW PRESET — BACKGROUND REPLACE: YOUR VIDEO, A NEW BACKGROUND, NOTHING ELSE</h3><p>“MiniMax H3 Background Replace — Pika H3 Flex 6 Steps — 261104”</p><p>Fully automatic: give it your video and a background image, and press Run. The person is found by itself and stays exactly as filmed: every pixel of the face, hair, clothes and hands, the lip movements, the timing and the sound. Everything around them becomes the place in your image and moves with your camera.</p><p>A talking man on a grey wall goes to a tropical beach, and a walking man in New York goes to a futuristic city:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cYsV7JOulMcRjNYH.jpeg" /></figure><p>There is nothing to write or set. Add your two files, press Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_YprtZ7XVYA9G8z4.png" /></figure><p>Long clips work too; we tested 31 seconds (736 frames). The video is made in overlapping windows, so a long clip needs no more GPU memory than a short one, only more time. The result is about 1 megapixel; run it through the Video Upscale preset below for a sharper video. People of similar size all stay, and distant passers-by go.</p><h3>NEW PRESET — VIDEO RELIGHT</h3><p>“MiniMax H3 Video Relight — Pika H3 Flex 6 Steps — 261104”</p><p>Relight a whole video. Add your clip and pick one of 20 looks such as neon noir, golden sunset backlight or blue hour, or design up to three lights yourself with their direction, softness and colour. The people, the motion, the camera and the set stay.</p><p>The same street clip in neon noir:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yNZ0i_cNEkhJfDht.jpeg" /></figure><p>The look, the lights and a preview sphere of where the key light comes from:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4wsBISDPdHH2fEn0.png" /></figure><p>H3 follows the plan through the prompt, so it is not pixel-exact light control; try two or three seeds.</p><h3>NEW PRESET — FACE SWAP</h3><p>“MiniMax H3 Face Swap — Ref2VA 8 Steps — 261104”</p><p>Put a new face on the person in your video. Add your clip and a clear close-up of the new face. Clothes, setting and camera stay as in your clip, and the body motion follows it.</p><p>A selfie clip with a new red-bearded face:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*X8ZgPIfklMD4mZyX.jpeg" /></figure><p>Two inputs and Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R0sjmDovUufIBtKL.png" /></figure><p>H3 makes the whole clip anew, so the timing of the motion can drift a little. For your clip’s exact timing and sound, use the Masked Edit preset below with target set to head.</p><h3>NEW PRESET — PIKA H3 FLEX CUT AND DRAG MOTION</h3><p>“Pika H3 Flex — Cut And Drag Motion To Video 6 Steps — 261104”</p><p>Draw a path on your photo and the subject follows it. Upload your photo and drag the path. The subject and the prompt are found and written for you. The subject is cut out and dragged along your path, and Pika H3 Flex turns that rough drag into a real video with moving light and sound.</p><p>From the photo, through the rough drag, to the finished drive:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*b53ag8LP4wp35J4r.jpeg" /></figure><p>Your photo, your path and Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2UyQuUH0Y82hwmc8.png" /></figure><p>Type a subject or a prompt whenever you want full control. Then the prompt model is not loaded at all.</p><h3>NEW PRESET — VIDEO UPSCALE</h3><p>“MiniMax H3 Video Upscale — Pika H3 Flex — 261104”</p><p>Fully automatic: give it your video and press Run. Your clip is encoded into MiniMax H3, enlarged by the H3 latent upscaler and refined by the last Pika H3 Flex steps, so the new detail is real video detail, not a sharpening filter. Your soundtrack stays exact.</p><p>Real footage reduced to 1280 x 704 and upscaled to 2560 x 1408, input on the left and result on the right, at the same moment:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*77xYupMDJ26CXAoh.jpeg" /></figure><p>Add your video and press Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9UcSAaKGan4rxdZD.png" /></figure><p>What to expect: landscapes, animals and textures gain the most. H3 redraws fine detail, so small text and fine logos can change; on our harbor test a boat’s small name came out garbled. One run makes the video at most twice as large.</p><h3>NEW PRESET — GENJUTSU MASKED EDIT: CHANGE ONE THING, KEEP EVERYTHING ELSE</h3><p>“MiniMax H3 Genjutsu — Masked Edit — Pika H3 Flex 6 Steps — 261104”</p><p>Replace a person or an object and keep the rest of your clip exactly as filmed. The street, the passers-by, the signs, the camera move and the sound stay untouched; only the target is regenerated and pasted back over your original frames.</p><p>A man on a New York street becomes a knight in armour, and the traffic behind him keeps moving:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hn4x89HkMFfldWVr.jpeg" /></figure><p>Your clip and the new character are all it needs. The person is found and the prompt is written for you:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GEJj5rGYAVg0pRt6.png" /></figure><p>To replace something else, name it in target: a car, a dog, a man in a blue suit. For a face swap with your clip’s exact timing and sound, set target to head and use a close-up of the new face.</p><h3>UPDATED — PIKA H3 FLEX: LONG VIDEO THAT TELLS A STORY</h3><p>“Pika H3 Flex — Text To Video 6 Steps — 261030”</p><p>A long video that develops instead of repeating. Switch on 9. LONG VIDEO, write your description, then a line with — -, then one scene, another — -, the next scene. Each part of the video follows its own scene.</p><p>Five scenes in one prompt: one continuous 20-second flight from a meadow to a sunset over the mountains:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GF_cxNc0mHH7Dask.jpeg" /></figure><p>The switch and the prompt:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*REvqydG-eyx60vif.png" /></figure><p>About 4 seconds per scene works well. Without — — the preset works as before: every window reads the same prompt, so motion can repeat. Scenes work in text runs, not with reference media.</p><h3>UPDATED — PROMPTS WRITE THEMSELVES</h3><p>“Pika H3 Flex — First Last Frame To Video 6 Steps — 261030”<br>“Pika H3 Flex — References To Video 6 Steps — 261030”</p><p>Leave the prompt empty and it is written for you. A local Qwen3-VL 8B looks at your frames or reference images and writes the MiniMax H3 prompt. Type your own whenever you want full control, and that model is not loaded at all.</p><h3>UPDATED — PIKA H3 FLEX: 2K</h3><p>Sharper video with one switch. In the Pika H3 Flex presets, turn on 8. 2K UPSCALE and the video is refined at 2560 pixels with the same sound. Same seed, without and with the 2K pass:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mjh1gYgxYAVOd4KS.jpeg" /></figure><p>Where to find it: in any of the three Pika H3 Flex presets, switch on 8. 2K UPSCALE, or 9. LONG VIDEO for clips over 15 seconds:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cRYo1lk4xVYJ_vC7.png" /></figure><p>Both switches start off, so your runs stay as before until you turn them on.</p><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V149.zip from this post and extract it over your installer folder.</li><li>Close ComfyUI, and SwarmUI too if it uses this ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install bundle 100.</li><li>Download the MiniMax-H3 Core Bundle with our SwarmUI Model Downloader V189 or newer.</li><li>Restart ComfyUI and refresh the browser tab.</li></ol><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>30 October 2026 Update V147</h3><p>Presets Folder Content In Latest Zip File</p><p>Pika H3 Flex: cinematic video with synchronized sound in 6 steps. Three new presets bring Pika Labs’ new Pika H3 Flex to ComfyUI: text to video, first + last frame to video and references to video, each with dialogue, sound effects and music made together with the picture.</p><p>Our three examples: a ship in a bottle that sails out onto a stormy ocean, a blacksmith whose golden eagle lands on his anvil, and an airship captain made from just two images:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4X5hnEC07_h9GDOj.jpeg" /></figure><p>Every result in this post is a real run of the shipped presets on our RTX 5090.</p><p>The complete Presets folder of V147: 116 ready-to-use ComfyUI presets plus 3 API presets, the new ones marked NEW:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1KAzTft0ui3dkJXe.png" /></figure><p>The presets cover Anima, AvatarForever, ERNIE Image, FastH3 Omni, Fibo 2.0, FLUX.1 Dev, Kontext and SRPO, FLUX 2, FLUX 3 Klein, Ideogram 4, Krea 2, LTX 2, LTX 2.3, LTX 2.5, MiniMax H3 with Genjutsu, MiniMax Music 3, MOSS-Image, Phantom Wan, Pika H3 Flex, Qwen Image, SDXL, Stable Video Infinity 2 Pro, Wan 2.1, Wan 2.2 and Z-Image.</p><h3>NEW PRESETS — PIKA H3 FLEX: VIDEO WITH SOUND IN 6 STEPS</h3><p>“Pika H3 Flex — Text To Video 6 Steps — 261030”<br>“Pika H3 Flex — First Last Frame To Video 6 Steps — 261030”<br>“Pika H3 Flex — References To Video 6 Steps — 261030”</p><p>Pika’s tuned MiniMax H3, in 6 steps. A 5-second 1344 × 768 clip with stereo sound takes about a minute on an RTX 5090. Each preset switches by itself between the text and frames model and the references model.</p><p>To run them, download the MiniMax-H3 Core Bundle. It now includes the two Pika H3 Flex LoRAs; if you already have the bundle, only those 3.6 GB download:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kX1KyXS-4K0DWWA8.png" /></figure><p>First + Last Frame: give it two images and it makes the shot between them. Our ship in a bottle becomes the real ship on the open sea in one continuous camera move:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iOHHytQEnJEXYMNk.jpeg" /></figure><p>Where to find it: Workflows → SECourses → Pika H3 Flex — First Last Frame To Video 6 Steps — 261030. Upload your first frame in INIT IMAGE and your last frame in LAST FRAME, describe what happens and press Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4EXlWmXwb_7ATDdC.png" /></figure><p>The timing line Pika H3 Flex expects is added for you with the real length of your clip. Leave LAST FRAME empty for first frame to video.</p><p>Text To Video: one prompt gives the whole scene, with the blacksmith’s line, the hammer, the hiss of steam and the eagle’s wingbeats in the soundtrack:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*w3inlDmOW0QHAbMZ.jpeg" /></figure><p>References To Video: our captain image and sky harbour image went in, a walking, talking scene came out, with his beard, goggles, green scarf and compass kept:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SaO8JR9b2RdG4OkJ.jpeg" /></figure><p>Prompt batches, .txt folder batches, Init Audio, an optional LoRA stack and face refinement are built in. The MiniMax H3 prompt guide now covers Pika H3 Flex, so ChatGPT, Claude or Gemini can write your prompts.</p><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V147.zip from this post and extract it over your installer folder.</li><li>Close ComfyUI, and SwarmUI too if it uses this ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install bundle 100.</li><li>Download the MiniMax-H3 Core Bundle with our SwarmUI Model Downloader V187 or newer.</li><li>Restart ComfyUI and refresh the browser tab.</li></ol><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>28 October 2026 Update V146</h3><p>Presets Folder Content In Latest Zip File</p><p>Anima character transfer, MOSS-Image layout control and stronger MiniMax H3 Genjutsu. Put one character into any scene with Anima, place and pose people with simple color and depth maps in MOSS-Image, decode FastH3 videos faster with LightVAE, and swap characters in your clips more convincingly with the new Genjutsu task adapters.</p><p>MOSS-Image Color + Depth Control turned the adapter author’s two-person example maps into photos, and MOSS Image Edit then added snow and autumn without touching the two men:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oTtesXWk9ytX_8F6.jpeg" /></figure><p>Every result in this post is a real run of the shipped presets on our RTX 5090.</p><p>The complete Presets folder of V146: 113 ready-to-use ComfyUI presets plus 3 API presets, new ones marked NEW and updated ones UPDATED:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4pI4_BumniiE8abA.png" /></figure><p>The presets cover Anima, AvatarForever, ERNIE Image, FastH3 Omni, Fibo 2.0, FLUX.1 Dev, Kontext and SRPO, FLUX 2, FLUX 3 Klein, Ideogram 4, Krea 2, LTX 2, LTX 2.3, LTX 2.5, MiniMax H3 with Genjutsu, MiniMax Music 3, MOSS-Image, Phantom Wan, Qwen Image, SDXL, Stable Video Infinity 2 Pro, Wan 2.1, Wan 2.2 and Z-Image.</p><h3>NEW PRESETS — ANIMA: TAKE YOUR CHARACTER INTO A NEW SCENE</h3><p>“Anima — Advanced Text To Image — 261028”<br>“Anima — Reference Character And Style — 261028”</p><p>One image of your character, any new scene. The Anima reference adapter carries the face, hair and outfit, and part of the drawing style, from your image, while your prompt picks the pose, place and light. Every final image is a clean 2048 × 2048.</p><p>To run them, download the Anima Core Bundle (3.84 GB) from the SwarmUI Bundles tab of our model downloader:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*F1hM6QOQZjFvlDNk.png" /></figure><p>Our example: one character made with text to image, then four new scenes from that single image in one prompt batch:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9H5HZZlOdBtL08U0.jpeg" /></figure><p>Where to find it: Workflows → SECourses → Anima — Reference Character And Style — 261028. Add your character with + Add references, write the new scene, add more scenes with + Prompt and press Run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zWx3KecgKbVZ1CqG.png" /></figure><p>Prompt batches, .txt folder batches and an optional LoRA stack are built in. The new Anima_Enchance_Prompt_Feed_For_LLMs.txt lets ChatGPT, Claude or Gemini write your prompts or a whole scene series.</p><h3>NEW PRESETS — MOSS-IMAGE COLOR + DEPTH CONTROL: PLACE AND POSE PEOPLE WITH SIMPLE MAPS</h3><p>“MOSS Image — Color + Depth Control — 261027”<br>“MOSS Image — Turbo 8 Steps Color + Depth Control — 261027”</p><p>Draw the layout, MOSS-Image follows it. A color map marks where each person or object goes, a depth map shows what is near and far, and your prompt says who they are. The Turbo preset makes a 1024 × 1024 image in about 7 seconds.</p><p>To run them, download the MOSS-Image Core Bundle. It now includes the color + depth adapter; if you already have the bundle, only that 0.29 GB file downloads:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MYcvAz6a6i_2CgFV.png" /></figure><p>The 50-step preset after the run that made the park photo at the top of this post:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Y7Hc2mt64Y_1vVCx.png" /></figure><p>Load your maps with choose file to upload in each map node. Part 8 of the MOSS-Image prompt guide writes the prompt for your maps.</p><h3>UPDATED PRESETS — MOSS IMAGE EDIT: NOW ON COMFYUI’S NATIVE EDIT SUPPORT</h3><p>“MOSS Image Edit — Advanced Image Edit — 261022”<br>“MOSS Image Edit — Turbo 8 Steps — 261022”</p><p>Both edit presets now run on ComfyUI’s own MOSS-Image edit support and open with an empty gallery: add your photo, type the instruction, press Run. Up to 4 images, and the Turbo edit takes about 6 seconds. They use the same MOSS-Image Core Bundle.</p><h3>NEW PRESET — FASTH3 OMNI REFERENCES TO VIDEO WITH THE LIGHTVAE DECODER</h3><p>“FastH3 Omni — LightVAE References To Video 8 Steps — 261027”</p><p>The same FastH3 Omni video, decoded faster. LightVAE, a lighter video decoder from LightX2V, decoded and saved our 5-second video in under 6 seconds instead of over 8, with nearly the same picture.</p><p>To run it, download the FastH3 Omni Core Bundle. It now includes the LightVAE decoder and the Genjutsu adapters:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fHQ38rRK7-N926Uc.png" /></figure><p>The preset after a real run:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_IImF6TWuAPx5Kw0.png" /></figure><p>Switch the decoder node to Original video VAE and press Run to compare: ComfyUI only decodes again.</p><h3>UPDATED PRESETS — MINIMAX H3 GENJUTSU: CHARACTER SWAP AND MOTION TRANSFER ADAPTERS</h3><p>“MiniMax H3 Genjutsu — FastH3 Omni 8 Steps — 261024”<br>“MiniMax H3 Genjutsu — Ref2VA 20 Steps — 261024”</p><p>Swap the person in any clip for your character, now with adapters trained for exactly this job. Genjutsu keeps the clip’s motion, timing, camera and sound; the new Character Swap and Motion Transfer adapters by Alisson Anjos make the transfer much stronger. Clips with mono sound now work too.</p><p>The FastH3 Omni 8 Steps preset uses the FastH3 Omni Core Bundle above. The Ref2VA 20 Steps preset uses the MiniMax-H3 Core Bundle, which now includes the three Genjutsu adapters:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R-LxhNwmzW3wJzbC.png" /></figure><p>Our armoured warrior in a shadowboxing clip, made by both updated presets:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7E3RxxM4SZVaCQVy.jpeg" /></figure><p>The FastH3 Omni Genjutsu preset after that run, with the new parts numbered:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1JuQHyECXF1xI7me.png" /></figure><p>Pick video background to keep the clip’s scene or image to keep your character’s background. Low VRAM mode is on by default with the same output.</p><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V146.zip from this post and extract it over your installer folder.</li><li>Close ComfyUI, and SwarmUI too if it uses this ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install bundle 100.</li><li>Download the core bundles shown above with our SwarmUI Model Downloader V186 or newer: Anima, MOSS-Image, FastH3 Omni or MiniMax-H3.</li><li>Restart ComfyUI and refresh the browser tab.</li></ol><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=978de93669a2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SwarmUI Full Tutorial: 1-Click Install, 4 MP Images, Full HD Video with Sound, Lip Sync & Edits]]></title>
            <link>https://medium.com/@furkangozukara/swarmui-full-tutorial-1-click-install-4-mp-images-full-hd-video-with-sound-lip-sync-edits-fef73fc4d600?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/fef73fc4d600</guid>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 01:38:15 GMT</pubDate>
            <atom:updated>2026-11-04T01:38:15.539Z</atom:updated>
            <content:encoded><![CDATA[<h3>Full tutorial link &gt; <a href="https://www.youtube.com/watch?v=0HmIeeI1iBo">https://www.youtube.com/watch?v=0HmIeeI1iBo</a></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WWTzsRpab9tv21Zf" /></figure><p>youtube.com</p><p>SwarmUI, set up with our one-click installers, our model downloader (35+ bundles, 517+ models) and 86+ ready presets, turns your own PC into a complete AI studio: 4-megapixel images, one-prompt edits, full HD MiniMax H3 videos with speech, sound effects and music, lip sync and Genjutsu person swaps. This tutorial installs everything from zero on Windows and teaches every feature step by step; the same installers run on RunPod, SimplePod, Massed Compute and Linux.</p><p>Links:</p><p>Download the SwarmUI installer, model downloader and presets (our SwarmUI post): [ <a href="https://www.patreon.com/posts/swarmui-auto-and-114517862">https://www.patreon.com/posts/swarmui-auto-and-114517862</a> ]</p><p>ComfyUI backend installer (our ComfyUI post): [ <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a> ]</p><p>One-time Windows requirements installer tutorial: [ <a href="https://youtu.be/ptuLL0C72Ms">https://youtu.be/ptuLL0C72Ms</a> ]</p><p>SwarmUI + ComfyUI on cloud GPUs (RunPod, Massed Compute): [ <a href="https://youtu.be/bBxgtVD3ek4">https://youtu.be/bBxgtVD3ek4</a> ]</p><p>SimplePod tutorial: [ <a href="https://youtu.be/yOj9PYq3XYM">https://youtu.be/yOj9PYq3XYM</a> ]</p><p>SECourses Discord: [ <a href="https://discord.com/invite/software-engineering-courses-secourses-772774097734074388">https://discord.com/invite/software-engineering-courses-secourses-772774097734074388</a> ]</p><p>You will learn the full route on a fresh Windows PC: download and extract both packages, install our ComfyUI backend and SwarmUI, connect them, import our presets, download models in one-click bundles and update everything, then every image, edit, video, lip-sync, audio and tool feature listed in the chapters below.</p><p>Main topics include MiniMax H3 and FastH3, Krea 2 Turbo, Qwen Image 2.1, FLUX 3 Klein, FLUX Fill and Kontext, Ideogram 4, Z Image, SDXL, LTX 2.3, SVI 2 Pro, Genjutsu, Long Lip Sync, RefMods and our LLM prompt kits. Every result was made on the fresh install with an RTX 5090.</p><p>Use the chapters to jump to any install step, feature, setting or comparison.</p><p>Chapters:</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=0">00:00:00</a> One prompt: H3 video with speech, effects and music</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=64">00:01:04</a> What our installers, downloader and presets add</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=143">00:02:23</a> Quick first image at 4 MP with Krea 2 Turbo</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=216">00:03:36</a> Download both zips from our posts</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=376">00:06:16</a> Create the folders and extract both packages</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=465">00:07:45</a> Install our ComfyUI backend: Torch 2.13, CUDA 13</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=550">00:09:10</a> Install SwarmUI with one double-click, first-run setup</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=692">00:11:32</a> Connect the ComfyUI backend and its arguments</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=802">00:13:22</a> Import our 86 presets with one bat file</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=850">00:14:10</a> The routine: reset, then Direct Apply a preset</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=909">00:15:09</a> Model downloader: the Krea 2 bundle in one click</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=1104">00:18:24</a> More bundles, single models, URL downloads</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=1238">00:20:38</a> First 4 MP image, bundle cards and model report</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=1385">00:23:05</a> Update in six steps, verified models kept</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=1678">00:27:58</a> What one preset click sets, our 11 extensions</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=1790">00:29:50</a> Tour of the Generate page</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=1963">00:32:43</a> Find the right preset for each job</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2075">00:34:35</a> Text in images: Ideogram 4 vs Qwen Image 2.1</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2219">00:36:59</a> Size, aspect ratio and our Megapixels option</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2289">00:38:09</a> Images, seeds and Variation Seed</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2396">00:39:56</a> Steps 4, 8, 16 compared; CFG, sampler, Sigma Shift</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2507">00:41:47</a> SDXL: negative prompt and CFG 3, 7, 12 compared</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2648">00:44:08</a> Z Image seed variety trick, Reuse Parameters</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2770">00:46:10</a> Prompt weights, alternate and from-to syntax</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2914">00:48:34</a> Regional prompt: a white wolf in one area</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=2958">00:49:18</a> Random picks, wildcards and variables</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=3141">00:52:21</a> Auto segment faces, more tags, LLM prompt kits</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=3276">00:54:36</a> LoRAs: Skin Fix at weight 0.6 vs 1.0</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=3426">00:57:06</a> FLUX Kontext edit and fixing model metadata</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=3581">00:59:41</a> Refine and upscale a draft to 4 MP</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=3752">01:02:32</a> Init Image creativity 0.3, 0.55 and 0.8</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=3881">01:04:41</a> Inpaint a dragon: Image Editor and SAM 2 masks</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=4055">01:07:35</a> Remove a watermark, every other pixel kept</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=4212">01:10:12</a> Outpainting in small steps with FLUX Fill</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=4433">01:13:53</a> Qwen Image 2.1: gold armor from one prompt</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=4583">01:16:23</a> Combine images; FLUX 3 Klein vs Qwen edit</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=4726">01:18:46</a> Layout boxes place every object exactly</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=4906">01:21:46</a> ControlNet Canny and Redux style transfer</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=5172">01:26:12</a> Attach media, trim clips, prompt batches</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=5343">01:29:03</a> H3 references, @ tokens and the token meter</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=5448">01:30:48</a> Write an H3 prompt with our kit and Gemini</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=5512">01:31:52</a> Text to video with sound: MiniMax H3 at 1080p</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=5615">01:33:35</a> 8-step turbo LoRA vs 20 steps, RIFE smoothing</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=5774">01:36:14</a> Image to video: the portrait moves and speaks</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=5862">01:37:42</a> First and last frame: black armor to gold</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=5942">01:39:02</a> Video to video with LTX 2.3: anime style</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=6015">01:40:15</a> Genjutsu: swap the person, keep every move</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=6192">01:43:12</a> Longer videos: continue a clip, SVI 2 Pro</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=6381">01:46:21</a> Long Lip Sync: a portrait speaks in full HD</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=6579">01:49:39</a> References to video: face, dragon and voice</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=6675">01:51:15</a> Save a character and voice as a RefMod</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=6824">01:53:44</a> Sound only with H3, Foley for a silent clip</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=7039">01:57:19</a> Grid Generator: samplers by steps</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=7191">01:59:51</a> Compare results, History and the video editor</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=7338">02:02:18</a> The ComfyUI workflow behind a preset</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=7456">02:04:16</a> Speed overrides and a second GPU backend</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=7564">02:06:04</a> Server, user settings, Utilities and the API</p><p><a href="https://youtu.be/0HmIeeI1iBo?t=7760">02:09:20</a> Recap and where to download everything</p><h3>4 November 2026 Update V189</h3><p>Edit real videos with MiniMax H3 in SwarmUI. Five new presets: replace the background of any video from one image, relight a whole clip, put a new face on someone, replace a person or object while every other pixel stays as filmed, and make long Pika H3 Flex videos that follow a story, scene by scene.</p><p>Amazing SwarmUI Presets moves to V88 with 99 presets: 5 new since V86, every earlier preset unchanged.</p><p>Real footage and real runs of the shipped presets:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*EiQBGZK4q6Rd2ON2.jpeg" /></figure><p>Every result in this post comes from a real run of the shipped presets in SwarmUI.</p><p>All 99 presets in the Presets tab, the new ones marked NEW:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iB8JvnntQxFb95IS.jpeg" /></figure><p>The new presets up close, each with its own animal thumbnail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YMeAkaxI3qZK3nCa.png" /></figure><p>Windows_Preset_Delete_Import.bat brings all of them into your SwarmUI; the steps are in How to update at the end.</p><p>To run them, download the MiniMax-H3 Core Bundle. It includes the FaceSwap LoRA; if you already have the bundle, only the missing files download:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*XLQFMYPor1tgolJw.png" /></figure><p>All the new presets use this one bundle.</p><p>Our live token counter is above the prompt: it counts your text, references, audio and video as you work and shows them against the MiniMax H3 limit, in green, amber and red, so you know what a run needs before you click Generate.</p><h3>NEW PRESET — BACKGROUND REPLACE: YOUR VIDEO, A NEW BACKGROUND, NOTHING ELSE</h3><p><em>“MiniMax H3 Background Replace — Pika H3 Flex 6 Steps — 261104”</em></p><p>Fully automatic: attach your video and a background image, and click Generate. The person is found by itself and stays exactly as filmed: every pixel of the face, hair, clothes and hands, the lip movements, the timing and the sound. Everything around them becomes the place in your image and moves with your camera.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yC6p6IQsGUx7qy_7.jpeg" /></figure><p>Where to find it: Presets tab → filter 261104 → apply the preset, attach both files beside the prompt and click Generate. There is nothing to write or set. Long clips work too: the video is made in overlapping windows, so a long clip needs no more GPU memory, only more time. The result is about 1 megapixel. People of similar size all stay, and distant passers-by go.</p><h3>NEW PRESET — VIDEO RELIGHT</h3><p><em>“MiniMax H3 Video Relight — Pika H3 Flex 6 Steps — 261104”</em></p><p>Relight a whole video. Attach your clip and click Generate: the prompt already holds a complete neon noir lighting plan. The people, the motion, the camera and the set stay.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TsIrn_ZE-tCUDDnM.jpeg" /></figure><p>Where to find it: Presets tab → filter 261104. Change the colours or directions in the Lighting plan lines of the prompt. H3 follows the plan through the prompt, so it is not pixel-exact light control; try two or three seeds.</p><h3>NEW PRESET — FACE SWAP</h3><p><em>“MiniMax H3 Face Swap — Ref2VA 8 Steps — 261104”</em></p><p>Put a new face on the person in your video. Attach your clip and a clear close-up of the new face, leave the prompt empty and click Generate. Clothes, setting and camera stay as in your clip.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RC3euSwidALYlv3c.jpeg" /></figure><p>H3 makes the whole clip anew, so the timing of the motion can drift a little. For your clip’s exact timing and sound, use the Masked Edit preset below with Genjutsu Target set to head.</p><h3>NEW PRESET — GENJUTSU MASKED EDIT: CHANGE ONE THING, KEEP EVERYTHING ELSE</h3><p><em>“MiniMax H3 Genjutsu — Masked Edit — Pika H3 Flex 6 Steps — 261104”</em></p><p>Replace a person or an object and keep the rest of your clip exactly as filmed. Attach your clip and the new character and click Generate: the person is found and the prompt is written for you. The street, the traffic, the signs and the camera move stay untouched.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*e7Jgo_E2A4GL9S5j.jpeg" /></figure><p>Where to find it: Presets tab → filter 261104. To replace something else, name it in Genjutsu Target: a car, a dog, a man in a blue suit.</p><h3>NEW PRESET — PIKA H3 FLEX LONG VIDEO THAT TELLS A STORY</h3><p><em>“Pika H3 Flex — Long Text to Video 20 Seconds — 261104”</em></p><p>A long video that develops instead of repeating. Write your description, then a line with — -, then one scene, another — -, the next scene. Each part of the video follows its own scene.</p><p>Five scenes in one prompt, one continuous 20-second flight:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UG_avIwTpkCT3q5Z.jpeg" /></figure><p>Where to find it: Presets tab → filter 261104. About 4 seconds per scene works well; set Text2Video Frames for other lengths (481 = 20 s, 719 = 30 s). Without — — every window reads the same prompt, so motion can repeat. Scenes work in text runs, not with reference media.</p><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v189.zip from this post and extract it over your downloader folder.</li><li>Close SwarmUI and run Windows_Install_or_Update_SwarmUI.bat.</li><li>Update the ComfyUI backend. With our ComfyUI installation, close SwarmUI again (step 2 starts it) and run Windows_Install_Or_Update_ComfyUI.bat from our ComfyUI V149. With SwarmUI’s own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start.</li><li>Restart SwarmUI, then run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v88.json. It backs up your current presets first, then replaces them all, your own presets too.</li><li>Run Windows_Start_Download_Models_App.bat and download the MiniMax-H3 Core Bundle.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>30 October 2026 Update V187</h3><p>Pika H3 Flex: cinematic video with synchronized sound in 6 steps. Three new presets bring Pika Labs’ new Pika H3 Flex to SwarmUI: text to video, beginning and ending frame to video and references to video, each with dialogue, sound effects and music made together with the picture.</p><p>Amazing SwarmUI Presets moves to V86 with 94 presets: 3 new, every earlier preset unchanged.</p><p>Our three examples: a ship in a bottle that sails out onto a stormy ocean, a blacksmith whose golden eagle lands on his anvil, and an airship captain made from just two images:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NlRLKQQ9n0edSAyr.jpeg" /></figure><p>Every result in this post is a real run of the shipped presets in SwarmUI on our RTX 5090.</p><p>All 94 presets of Amazing SwarmUI Presets V86 in SwarmUI’s Presets tab, the new ones marked NEW:</p><p>The 3 new presets up close, each with its own animal thumbnail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DZLA0s0UBxjXd_mo.png" /></figure><p>Windows_Preset_Delete_Import.bat brings all of them into your SwarmUI; the steps are in How to update at the end.</p><h3>NEW PRESETS — PIKA H3 FLEX: VIDEO WITH SOUND IN 6 STEPS</h3><p><em>“Pika H3 Flex — Text to Video 6 Steps — 261030”</em></p><p><em>“Pika H3 Flex — Beginning and Ending Frame to Video 6 Steps — 261030”</em></p><p><em>“Pika H3 Flex — References to Video 6 Steps — 261030”</em></p><p>Pika’s tuned MiniMax H3, in 6 steps. A 5-second 1344 × 768 clip with stereo sound takes about a minute on an RTX 5090. Init Audio, face inpainting and your own MiniMax H3 LoRAs stay available, as in every MiniMax H3 preset.</p><p>To run them, download the MiniMax-H3 Core Bundle. It now includes the two Pika H3 Flex LoRAs; if you already have the bundle, only those 3.6 GB download:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W3COmHeANdTqaYZk.png" /></figure><p>Beginning and Ending Frame: give it two images and it makes the shot between them. Our ship in a bottle becomes the real ship on the open sea in one continuous camera move:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HF7EjaYObTGmI_fd.jpeg" /></figure><p>Where to find it: Presets tab → filter Pika → Direct Apply Pika H3 Flex — Beginning and Ending Frame to Video 6 Steps — 261030. Put the beginning frame in Init Image, tick Display Advanced Options and put the ending frame in Image To Video → Video End Image, then click Generate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*32WYrXuPwykQV1Cd.png" /></figure><p>Start the prompt with Pika’s timing line; the preset description has it ready to copy. Leave Video End Image empty for first frame to video.</p><p>Text to Video: one prompt gives the whole scene, with the blacksmith’s line, the hammer, the hiss of steam and the eagle’s wingbeats in the soundtrack:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Q1b_WvXQxLKlbbTV.jpeg" /></figure><p>References to Video: attach a character image and a place image beside the prompt; our captain walks his sky harbour with his beard, goggles, green scarf and compass kept:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jfkH8-J5sUaUiOC3.jpeg" /></figure><p>The MiniMax H3 prompt guide now covers Pika H3 Flex, so ChatGPT, Claude or Gemini can write your prompts.</p><h3>ALSO IN V187</h3><ul><li>The MiniMax-H3 Low VRAM Bundle includes the two Pika H3 Flex LoRAs too.</li><li>SwarmUI’s scheduler list gains pika_h3_flex, Pika’s official 6-step schedule, after the update below.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v187.zip from this post and extract it over your downloader folder.</li><li>Close SwarmUI and run Windows_Install_or_Update_SwarmUI.bat.</li><li>Update the ComfyUI backend. With our ComfyUI installation, close SwarmUI again (step 2 starts it) and run Windows_Install_Or_Update_ComfyUI.bat from our ComfyUI V147. With SwarmUI’s own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start.</li><li>Restart SwarmUI, then run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v86.json. It backs up your current presets first, then replaces them all, your own presets too.</li><li>Run Windows_Start_Download_Models_App.bat and download the MiniMax-H3 Core Bundle.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>28 October 2026 Update V186</h3><p>Anima character transfer, MOSS-Image layout control, stronger MiniMax H3 Genjutsu and faster FastH3 decoding. Put one character into any scene with Anima, place and pose people with simple color and depth maps in MOSS-Image, swap characters in your clips more convincingly with the new Genjutsu task LoRAs, and decode FastH3 Omni videos faster with LightVAE.</p><p>Amazing SwarmUI Presets moves to V85 with 91 presets: 5 new, 2 updated, and 24 others now switch the new options off when you apply them.</p><p>With the adapter’s example maps, both new MOSS-Image presets put the two men where the mannequins stand, in their poses; without maps, the same prompt and seed picks its own layout:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Kx4fURofxT3W_t6D.jpeg" /></figure><p>Every result in this post is a real run of the shipped presets in SwarmUI on our RTX 5090.</p><p>All 91 presets of Amazing SwarmUI Presets V85 in SwarmUI’s Presets tab, the new ones marked NEW and the updated ones UPDATED:</p><p>The 5 new and 2 updated presets up close:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TTDk8Kf-6wMXyyMW.png" /></figure><p>Windows_Preset_Delete_Import.bat brings all of them into your SwarmUI; the steps are in How to update at the end.</p><h3>NEW PRESETS — ANIMA: TAKE YOUR CHARACTER INTO A NEW SCENE</h3><p><em>“Anima — Advanced Text To Image — 261028”</em></p><p><em>“Anima — Reference Character And Style — 261028”</em></p><p>One image of your character, any new scene. The Anima reference adapter carries the face, hair and outfit, and part of the drawing style, from your image, while your prompt picks the pose, place and light. Every final image is a clean 2048 × 2048.</p><p>To run them, download the Anima Core Bundle (3.84 GB) from the model downloader:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lYREYAiVf6m3VyAG.png" /></figure><p>Our example: one character made with text to image, then four new scenes from that single image with one click on Generate prompt batch:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VWA1yqGj6csXhQxO.jpeg" /></figure><p>Where to find it: Presets tab → filter Anima → Direct Apply Anima — Reference Character And Style — 261028. Upload your character in Anima Reference Image, add your scene prompts and click Generate prompt batch:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8jmYwTrUY8r3dwh1.png" /></figure><p>The new Anima_Enchance_Prompt_Feed_For_LLMs.txt lets ChatGPT, Claude or Gemini write your prompts or a whole scene series.</p><h3>NEW PRESETS — MOSS IMAGE COLOR + DEPTH CONTROL: PLACE YOUR PEOPLE WITH A MAP</h3><p><em>“MOSS Image — Color + Depth Control — 261027”</em></p><p><em>“MOSS Image — Turbo 8 Steps Color + Depth Control — 261027”</em></p><p>Draw the layout, MOSS-Image follows it. A color map marks where each person or object goes, a depth map shows what is near and far, and your prompt says who they are. Turbo makes a 1024 × 1024 image in about 4 seconds.</p><p>To run them, download the MOSS-Image Core Bundle. It now includes the color + depth adapter; if you already have the bundle, only that 0.29 GB file downloads:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rVtk-0VxIwAnbhFG.png" /></figure><p>Where to find it: Quick Tools → Reset Params to Default, then Presets tab → filter Color + Depth → Apply. Add your maps with the Upload buttons of MOSS Image Color Map and MOSS Image Depth Map, describe each person and press Generate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YeEl_dhSLSJjWBPa.png" /></figure><p>Part 8 of the MOSS-Image prompt guide writes the prompt for your maps.</p><h3>NEW PRESET — FASTH3 OMNI WITH THE LIGHTVAE VIDEO DECODER</h3><p><em>“FastH3 Omni — LightVAE References to Video 8 Steps — 261027”</em></p><p>The same FastH3 Omni video, decoded faster. LightVAE by LightX2V decoded and saved our 5-second video in about 4 seconds instead of about 6.5, with frames that look the same.</p><p>To run it, download the FastH3 Omni Core Bundle. It now includes the LightVAE decoder and the three Genjutsu files:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*a5JOw-SQLmmBXofP.png" /></figure><p>Where to find it: Quick Tools → Reset Params to Default, then Presets tab → filter LightVAE → Apply. The new MiniMax H3 Video Decoder parameter in Core Parameters picks the decoder for any MiniMax H3 or FastH3 Omni model:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RAVdrPNaQ3yxKujD.png" /></figure><p>Set it to None for the original decoder; every other H3 preset does that for you.</p><h3>UPDATED PRESETS — MINIMAX H3 GENJUTSU: CHARACTER SWAP AND MOTION TRANSFER LORAS</h3><p><em>“MiniMax H3 Genjutsu — FastH3 Omni 8 Steps — 261024”</em></p><p><em>“MiniMax H3 Genjutsu — Ref2VA 20 Steps — 261024”</em></p><p>Swap the person in any clip for your character, now with task LoRAs trained for exactly this job. Character Swap v1.1 and Motion Transfer v1.0 by Alisson Anjos bring your character’s look over much more closely, and clips with mono sound now work.</p><p>The FastH3 Omni 8 Steps preset uses the FastH3 Omni Core Bundle above. The Ref2VA 20 Steps preset uses the MiniMax-H3 Core Bundle, which now includes the three Genjutsu files:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TenO5UeZw6JYkH4Q.png" /></figure><p>Same warrior, same shadowboxing clip, same seed: the original method and the new task LoRA default:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MrVwek3zYakvhQKQ.jpeg" /></figure><p>The task LoRA brings his big scale shoulder guards and heavy build over. original Genjutsu is still one click away in Genjutsu Method.</p><p>Where to find it: Quick Tools → Reset Params to Default, then Presets tab → filter H3 Genjutsu → Apply. Attach the clip and your character with Add References and press Generate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xXPkKCTH3-DJDInT.png" /></figure><p>Low VRAM mode is on by default with the same output.</p><h3>ALSO IN V186</h3><ul><li>The MiniMax-H3 Low VRAM Bundle and the Complete Image Generation and Editing Bundle include the new models too.</li><li>The Krea 2 and FLUX 3 Klein bundles no longer download their shared text encoder again on every run.</li><li>MiniMax H3 Long Lip Sync — References no longer adds a second pass after an image-to-video preset: about 5.5 minutes instead of 11.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v186.zip from this post and extract it over your downloader folder.</li><li>Close SwarmUI and run Windows_Install_or_Update_SwarmUI.bat.</li><li>Update the ComfyUI backend. With our ComfyUI installation, close SwarmUI again (step 2 starts it) and run Windows_Install_Or_Update_ComfyUI.bat from our ComfyUI V146. With SwarmUI’s own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start.</li><li>Restart SwarmUI, then run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v85.json. It backs up your current presets first, then replaces them all, your own presets too.</li><li>Run Windows_Start_Download_Models_App.bat and download the core bundles shown above.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>24 October 2026 Update V185</h3><p>This is a MiniMax H3 Genjutsu update. V185 adds two presets that replace the person in any video with your own character: attach your clip and one image of the new character, press Generate, and the result keeps the clip’s motion, timing, camera, length, aspect ratio and original sound. Our MiniMaxH3References extension gets the new MiniMax H3 Genjutsu group, and the model downloader adds SAM 3.1 and Qwen3-VL 8B to the MiniMax H3 and FastH3 Omni bundles.</p><p>Amazing SwarmUI Presets moves to V84 with 86 presets: the two new presets, and the other 84 presets unchanged.</p><p>Our test in SwarmUI: this armoured warrior image and a 9.8-second phone clip of a man shadowboxing in a gym. Both presets put the warrior into the clip, punch for punch and step for step, in the clip’s portrait format and with its own sound:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wyGWFWdWS1Dd_d-K.jpeg" /></figure><p>Exact V184 → V185 zip audit:</p><ul><li>Presets: Amazing_SwarmUI_Presets_v83.json was replaced by Amazing_SwarmUI_Presets_v84.json. Its 84 presets are unchanged in every setting, description and thumbnail; 2 presets are new.</li><li>6 changed files: the model catalog and its size data (SAM 3.1 and Qwen3-VL 8B in the MiniMax-H3 Core, MiniMax-H3 Low VRAM and FastH3 Omni Core bundles), update_swarmui.py (a premium extension that could not be replaced during an update is installed again by the next update), the downloader app (shows v185), the regenerated Which_Bundles_Downloads_Which_Preset_Models.html and the MiniMax H3 prompt guide (new Genjutsu section).</li></ul><h3>NEW PRESETS — MINIMAX H3 GENJUTSU: PUT YOUR CHARACTER INTO ANY VIDEO</h3><p>“MiniMax H3 Genjutsu — FastH3 Omni 8 Steps — 261024”<br>“MiniMax H3 Genjutsu — Ref2VA 20 Steps — 261024”</p><p>How it works: SAM 3.1 finds the person in every frame of your clip, and only that person is colour-inverted. MiniMax H3 then copies the clip’s motion, timing, camera and lip movements, but not the performer’s face, hair or clothes. A local Qwen3-VL 8B (Auto Describe) looks at your character image and six frames of the clip and writes the prompt for you: the new outfit, the action, the setting, the lighting and the framing. The clip’s soundtrack is locked into MiniMax H3’s audio stream for lip sync, and the saved video carries the original sound.</p><p>Where to find it: Quick Tools → Reset Params to Default, then the Presets tab → type Genjutsu in the filter → the preset card’s menu → Direct Apply. Attach the clip and the character image with Add References (or Add A Reference With Trim to pick the part of the clip) and press Generate. Here is the FastH3 Omni preset after a real run of our example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Aoay9yXpffb4byV9.png" /></figure><ul><li>MiniMax H3 Genjutsu (on): turns the first reference video into the SAM 3.1 motion guide.</li><li>Genjutsu Background: video (character replacement) keeps the clip’s location, lighting and framing; image (motion transfer) keeps your character image’s background and takes only the motion.</li><li>Genjutsu Target: who to replace, in plain words: person, or man in red:1 to pick one of several people.</li><li>Genjutsu Auto Describe (on): the local Qwen3-VL 8B writes the outfit, action, setting, lighting and framing into the prompt. Leave the prompt empty, type short notes, or paste a complete MiniMax H3 prompt from ChatGPT, Claude or Gemini; a complete prompt is used as written.</li><li>Genjutsu Spoken Words (optional): the words spoken in the clip, for word-by-word lip sync.</li><li>Genjutsu Lock Clip Audio (on) and Genjutsu Match Clip Aspect (on): the clip’s sound drives the lip movements, and the output takes the clip’s aspect ratio at your Width x Height area.</li><li>Genjutsu Motion Guide Megapixels (0.15): the guide size; advanced options set the detection threshold and the mask grow.</li><li>The output follows the clip: our 9.8-second 1080x1920 clip gave 226 frames (9.4 seconds) at 768x1344, and the token meter beside the prompt shows that length and size before you press Generate.</li><li>Speed: FastH3 Omni 8 steps made the 9.4-second example in 4.5 minutes and Ref2VA 20 steps in 9.9 minutes in SwarmUI, first runs with model loading on an RTX 5090.</li></ul><h3>MODEL DOWNLOADER — SAM 3.1 AND QWEN3-VL 8B IN THE MINIMAX H3 BUNDLES</h3><p>The MiniMax-H3 Core Bundle, the MiniMax-H3 Low VRAM Bundle and the FastH3 Omni Core Bundle now download everything the Genjutsu presets use:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KrY29tMPWyKpMgCE.png" /></figure><ul><li>SAM 3.1 Multiplex FP16 (1.63 GB, saved to Models/Stable-Diffusion): finds the person to replace in every frame.</li><li>Qwen3-VL 8B INT8 ConvRot (8.71 GB, saved to Models/text_encoders): writes the Genjutsu prompt. It is the same file as the Qwen Image 2.1 encoder, so if you already have it, the downloader checks it and keeps it.</li><li>New totals: MiniMax-H3 Core Bundle 79.91 GB (19 files), MiniMax-H3 Low VRAM Bundle 64.21 GB (19 files), FastH3 Omni Core Bundle 60.12 GB (11 files), sizes as the downloader shows them.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v185.zip from this post and extract it over your existing downloader folder.</li><li>Close SwarmUI and run Windows_Install_or_Update_SwarmUI.bat. It installs MiniMaxH3References 1.22.1 and our FoleyExtension with the Genjutsu nodes.</li><li>Keep the ComfyUI backend up to date. If SwarmUI uses our ComfyUI installation, close SwarmUI again (step 2 starts it, and a running backend blocks the update) and update that ComfyUI with our ComfyUI installer V145 (Windows_Install_Or_Update_ComfyUI.bat), which also brings the same Genjutsu presets to ComfyUI. If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start.</li><li>Restart SwarmUI and its ComfyUI backend so the new extension and nodes register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v84.json. It backs up your current presets, thumbnails included, to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the MiniMax-H3 Core Bundle or the FastH3 Omni Core Bundle (the files you already have are checked and kept).</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>23 October 2026 Update V184</h3><p>This is a Fibo 2.0, MOSS Image Edit and FastH3 Omni update. V184 adds four presets for Bria’s Fibo 2.0, which generates and edits images from long JSON captions, two MOSS Image Edit presets that edit your photos with a short instruction, and four FastH3 Omni presets that turn your images, videos and sounds into video with synchronized audio, stills or audio in 8 steps. The model downloader gets the FastH3 Omni and Fibo 2.0 Core Bundles and a redesigned, much faster interface, keeps the model files whose header SwarmUI rewrote instead of downloading them again, and fixes the Qwen Image 2.1 text encoder that SwarmUI downloaded a second time.</p><p>Amazing SwarmUI Presets moves to V83 with 84 presets: 10 new presets, corrected descriptions for the four FLUX 2 8-step presets and the fixed text encoder in the two Qwen Image 2.1 Unified presets; the other 68 presets are unchanged. The new presets need the new FastH3 Omni Core Bundle, the Fibo 2.0 Core Bundle and the updated MOSS-Image Core Bundle of the model downloader.</p><p>All 84 presets as they appear in the SwarmUI Presets tab; the new presets are in the last rows:</p><p>Every preset still has its own animal. The ten new presets on their own:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*P92SbNXf73-2_WGH.png" /></figure><p>Four Fibo 2.0 presets (text to image and image editing, Dev and Turbo), two MOSS Image Edit presets (30 steps and Turbo 8 steps) and four FastH3 Omni presets (references to video, long lip sync, text to image and audio only), each with its own artwork.</p><p>Exact V183 → V184 zip audit:</p><ul><li>Presets: Amazing_SwarmUI_Presets_v81.json was replaced by Amazing_SwarmUI_Presets_v83.json. Of the 74 presets it keeps, 68 are unchanged in every setting, description and thumbnail; the four FLUX 2 8-step presets have a corrected description and the two Qwen Image 2.1 Unified presets select the fixed text encoder.</li><li>9 changed files: the model catalog and its size data (the FastH3 Omni and Fibo 2.0 models and bundles, the MOSS-Image edit LoRA, the Qwen Image 2.1 encoder names and the new source of the MiniMax H3 video VAE), the downloader app (new interface, shows v184), HF_model_downloader.py (keeps files whose header SwarmUI rewrote), update_swarmui.py (the new Fibo2 premium extension), the bundle report generator and the regenerated Which_Bundles_Downloads_Which_Preset_Models.html, and the MOSS-Image and MiniMax H3 prompt guides.</li><li>New: the Fibo 2.0 prompt guide in Automatic_Prompt_Generate_Enhance_To_Video_And_Image_Models_LLM_Instructions.</li><li>Windows_Install_or_Update_SwarmUI.bat brings SwarmUI with its native Fibo 2.0 support, the new Fibo2 premium extension, MossImage with Image edit mode, MiniMaxH3References 1.21.0, QwenImage21 1.0.2, SECoursesAudioTools and ComfyUI-TeaCache. The ComfyUI backend is updated separately (step 3 of HOW TO UPDATE).</li></ul><h3>NEW PRESETS — FIBO 2.0: TEXT TO IMAGE AND IMAGE EDITING, DEV AND TURBO</h3><p><em>“Fibo 2 Dev — Advanced Text To Image — 261022”</em></p><p><em>“Fibo 2 Dev — Advanced Image Edit — 261022”</em></p><p><em>“Fibo 2 Turbo — Image Edit 4 Steps — 261022”</em></p><p><em>“Fibo 2 Turbo — Text To Image 4 Steps — 261022”</em></p><p>Fibo 2.0 by Bria is an 8B image model that generates and edits images. It was trained on long structured JSON captions that describe every object, the lighting, the camera and the text in the image, so a JSON caption gives the most control; plain text also works. All four presets load the models in INT8 ConvRot with the Qwen3-VL 4B INT8 ConvRot text encoder and the FLUX 2 VAE, set 1024×1024 and turn on our new Fibo2 extension. Measured in SwarmUI on an RTX 5090 at 1024×1024, for a new prompt with the models already loaded:</p><ul><li>Dev — Advanced Text To Image: Bria’s official recipe: 50 Euler steps, CFG 5 and Bria’s JSON negative prompt from the model card. 17.3 seconds per image.</li><li>Turbo — Text To Image 4 Steps: Bria made the 4-step Edit Turbo model for editing, but in our tests it also renders text to image very well, lettering included: 4 steps, CFG 1, 1.9 seconds per image.</li><li>Dev — Advanced Image Edit: the official editing recipe with 50 steps and CFG 5: 34.7 seconds.</li><li>Turbo — Image Edit 4 Steps: Bria’s distilled editing model, 4 steps, CFG 1: 2.5 seconds.</li></ul><p>About 1 megapixel is the trained size (1024×1024, 1424×800, 1248×832 and the other shapes); 2 megapixels such as 1664×1248 also work well.</p><p>Where to find it: Presets tab → type Fibo 2 in the filter → click a preset card → Apply. Its settings fill the new Fibo 2.0 group of the parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UKAiXzPeecQoeDfl.png" /></figure><ul><li>Fibo 2 Advanced (on): encodes your prompt with ComfyUI’s native Fibo 2.0 text encoder exactly like Bria’s pipeline. SwarmUI’s generic encoder wraps Fibo 2.0 prompts differently and leaves edit images out of the text encoder.</li><li>Fibo 2 Sigma Shift: Bria’s official shift, which changes with the image size.</li><li>Fibo 2 Edit Size: edits keep the size of the first image from 1 to 4 megapixels.</li><li>Editing: attach 1 to 4 images to the prompt with Add Image / Video / Audio (or drag them onto the prompt box) and write the instruction; @image1, @image2 refer to them. Masked edit: an Init Image plus a mask painted in the image editor (white = change).</li><li>Prompt syntax: a JSON caption in Bria’s format (the new Fibo 2.0 LLM guide writes them for you) or plain text. {red|blue} picks one at random following the seed; (word:1.2) weights are removed because Qwen3-VL would read them as text.</li><li>Fibo 2 Prompt List: prompt batches from typed prompts, .txt files or a whole folder (one prompt per file, JSON captions included), like the MOSS-Image prompt batches. Generate prompt batch runs your main prompt first, then every extra prompt.</li><li>The extension acts only on Fibo 2.0 models. Apply any other preset and its own workflow runs unchanged.</li></ul><h3>NEW — ONE JSON PROMPT: FIBO 2 DEV 50 STEPS VS TURBO 4 STEPS</h3><p>The same JSON caption and seed (261023) through both text-to-image presets. The caption asks for a cafe on a rainy evening with a green sign that reads “MORNING BREW” and a sidewalk chalkboard with “Fresh Coffee” and “Open Late”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eGmSuGjMTo_POR9C.jpeg" /></figure><p>Both presets wrote every word as asked, with the warm window light, the wet cobblestones and the guests inside. Turbo needs 1.9 seconds instead of 17.3, which suits trying many prompts; Dev is the full, non-distilled model with CFG and Bria’s official recipe, and the model to train LoRAs on.</p><h3>NEW — ONE PORTRAIT, THREE NEW EDIT PRESETS</h3><p>A portrait made with Fibo 2 Turbo Text To Image, attached to the prompt as Image 1 and edited with each new edit preset and its default instruction (seed 261023):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Sd8YEI-dFo5dCuQl.jpeg" /></figure><p>Both Fibo 2.0 edits redrew the photo as an anime illustration and kept the pose, the clothes, the coffee cup and the flower shop. MOSS Image Edit replaced the street with a sunny beach and kept the same face, hair, clothes and pose, in 11.1 seconds.</p><h3>NEW PRESETS — MOSS IMAGE EDIT: EDIT YOUR PHOTOS WITH ONE SENTENCE</h3><p><em>“MOSS Image Edit — Advanced Image Edit — 261022”</em></p><p><em>“MOSS Image Edit — Turbo 8 Steps — 261022”</em></p><p>Both presets use MOSS-Image SFT INT8 ConvRot with ChenkinLab’s Moss-Image-Edit LoRA, which the MossImage extension’s new Image edit mode adds by itself together with the adapter’s reference branch. Attach up to 4 images: image 1 is edited and the others are references, named image 1, image 2 (or @image1, @image2) in the instruction. Say what changes and what stays, for example “Change the background to a sunny beach, preserving the person’s identity and pose.”</p><ul><li>Advanced Image Edit: the official edit recipe: 30 steps, CFG 4, shift 5 and an empty negative prompt. 11.1 seconds at 1024×1024 on an RTX 5090.</li><li>Turbo 8 Steps: the community 8-step DMD LoRA at 0.7 with 8 steps and CFG 1: 2.4 seconds. Turbo is close to the 30-step preset in many edits, but for our busy street portrait it kept the street on all five seeds we tried, while the 30-step preset gave a beach on all five. When Turbo ignores a big change such as a new background, use the 30-step preset.</li><li>The result has image 1’s shape at up to 1024 pixels on the long side (MOSS Image Edit Reference Size; 1536 also works).</li><li>Colors, clothes, expressions, lighting, seasons, backgrounds and style edit best. Placing an object from a second image works; copying its exact look is weaker.</li></ul><p>Where to find it: Presets tab → type MOSS Image Edit → click a preset card → Apply, then add the photo with Add Image / Video / Audio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*adA_GuOt_-JnrGJu.png" /></figure><p>The preset sets MOSS Image Mode to Image edit (attached images). The new MOSS_Image_Edit_LoRA.safetensors (0.28 GB) is part of the updated MOSS-Image Core Bundle.</p><h3>NEW PRESETS — FASTH3 OMNI: VIDEO WITH AUDIO FROM YOUR REFERENCES IN 8 STEPS</h3><p><em>“FastH3 Omni — References to Video 8 Steps — 261022”</em></p><p><em>“FastH3 Omni — Long Lip Sync — References 8 Steps — 261022”</em></p><p><em>“FastH3 Omni — Text to Image with Optional References — 261022”</em></p><p><em>“FastH3 Omni — Audio Only with Optional References — 261022”</em></p><p>FastH3 Omni by FastVideo and Nuva Lab is MiniMax H3’s references-to-video-and-audio model distilled to 8 steps. One model makes video with synchronized 32 kHz stereo audio from your prompt plus up to 9 images, 3 videos and 3 audio files, with no turbo LoRA: 8 steps, CFG 1. The presets load it in INT8 ConvRot with the INT8 ConvRot Qwen3-VL 32B text and vision encoder. The References to Video and Long Lip Sync presets also run the sparse attention FastH3 was trained with (VSA): sampling was about 1.4× faster for 5 seconds and 1.9× for 10 seconds of 1344×768 video on an RTX 5090, with the same quality in our tests.</p><ul><li>References to Video: 1344×768, 124 frames (5.2 seconds) with audio: 53 to 57 seconds in our runs with one reference image (one run took 108 seconds; we could not find why). It also works as text to video without references.</li><li>Long Lip Sync: put the complete soundtrack in Init Audio and a portrait beside the prompt as @image1; one 24 FPS video follows the whole audio in overlapping 8-step windows at 832×1248. About 3.5 minutes for 15 seconds of speech (3.32 minutes in our SwarmUI test).</li><li>Text to Image with Optional References: stills from text, or with up to 9 reference images: 9.8 seconds at 1344×768 for a new prompt with one reference image.</li><li>Audio Only with Optional References: 32 kHz stereo audio returned as a lossless FLAC: 8 seconds of audio in 7.2 seconds. Put spoken words inside &lt;d&gt;[English] … &lt;/d&gt;.</li></ul><p>Where to find it: first Quick Tools → Reset Params to Default, so no MiniMax H3 turbo LoRA from an earlier preset stays selected. Then Presets tab → type FastH3 Omni → click a preset card → Apply, add your images, videos or audio with Add References and mention them in the prompt as @image1, @video1, @audio1.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gh6LB2ruhEEEy6Y_.png" /></figure><p>The prompt describes the woman from @image1 in front of a flower shop: she smiles, raises her coffee cup, greets the viewer with a spoken line in &lt;d&gt;, laughs and takes a sip. Nine frames from the start to the end of the clip:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VPKTwHh6hKzYpErC.jpeg" /></figure><p>Her face, curly hair, freckles, earrings and clothes stay those of the portrait in every frame, and the clip follows the prompt from the greeting to the sip. Video and its 32 kHz stereo soundtrack came out of one 8-step run.</p><h3>NEW — MODEL DOWNLOADER: FASTH3 OMNI AND FIBO 2.0 CORE BUNDLES</h3><ul><li>FastH3 Omni Core Bundle (49.78 GB, 9 files): FastH3 Omni INT8 ConvRot (20.69 GB), the MiniMax H3 Qwen3-VL 32B INT8 ConvRot text and vision encoder (25.28 GB), the MiniMax H3 video VAE (2.62 GB) and audio VAE (0.56 GB), plus the face models the presets’ face refinement and mouth pass use. MiniMax H3 users already have the audio VAE and the face models, and the INT8 ConvRot encoder if they chose it. The video VAE now comes from Comfy-Org, the file SwarmUI itself downloads: if an earlier downloader saved Kijai’s copy at the same path, it is replaced once (2.62 GB).</li><li>Fibo 2.0 Core Bundle (20.50 GB, 4 files): Fibo 2.0 Dev and Edit Turbo in INT8 ConvRot (7.83 GB each), the Qwen3-VL 4B INT8 ConvRot text encoder (4.52 GB, saved as qwen3vl_4b_int8_convrot.safetensors so it never replaces the qwen3vl_4b.safetensors other models use) and the shared FLUX 2 VAE (0.31 GB).</li><li>MOSS-Image Core Bundle (21.91 GB, 7 files): now also downloads MOSS_Image_Edit_LoRA.safetensors (0.28 GB) for the MOSS Image Edit presets.</li></ul><p>Where to find it: run Windows_Start_Download_Models_App.bat. The FastH3 Omni Core Bundle and the Fibo 2.0 Core Bundle are the first two bundles of the SwarmUI Bundles tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*j0nAOor3VtOoc2Qb.png" /></figure><p>Optional, under Video Generation Models → FastH3 Omni Video Models: FastH3 Omni in BF16 (41.20 GB), FP8 (21.46 GB) and NVFP4 QAD (12.83 GB, for RTX 50 series and other Blackwell GPUs). Under Image Generation Models → Fibo 2.0 Models: both models in BF16 (15.32 GB each), the BF16 text encoder (8.27 GB, saved as qwen3vl_4b.safetensors, the name the FP8 encoder of FLUX 3 Klein and Krea 2 also uses, so keep one precision) and Comfy-Org’s Fibo 2.0 turbo LoRA (1.85 GB), which lets Dev and your Dev LoRAs run in 4 steps; the presets do not select it.</p><h3>NEW — THE DOWNLOADER KEEPS MODELS WHOSE HEADER SWARMUI REWROTE</h3><p>The first time SwarmUI uses a model, it writes the model hash into the file’s header (SwarmUI’s ImageMetadataIncludeModelHash setting, on by default). The tensor data stays the same, but the file gets a little larger or its bytes change, so V183 failed its size or SHA-256 check and downloaded the whole model again. V184 reads the original header from Hugging Face, checks your file’s tensor data behind it, and keeps the file when every tensor byte matches. A checked file is remembered, so the next run skips it at once.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*grub5ZoqVzKT6zw4.png" /></figure><p>The log says “[VERIFIED] … the tensor data matches the Hugging Face file; only the header metadata differs” and “[SKIP] … already complete”. On our test PC, 59 model files with 581 GB in total were a little larger than their Hugging Face files, the sign of SwarmUI’s header rewrite. The Fibo 2.0 Core Bundle’s four files were checked and kept in 25 seconds instead of downloading 20.50 GB again. Files whose tensor data differs are still downloaded again, as before.</p><h3>FIXED — SWARMUI NO LONGER DOWNLOADS THE QWEN IMAGE 2.1 TEXT ENCODER AGAIN</h3><p>Thank you to the supporter who reported this: after the Qwen Image 2.1 Core Bundle, SwarmUI said the text encoder was missing and downloaded it again. The downloader saved the INT8 ConvRot Qwen3-VL 8B encoder as qwen3vl_8b.safetensors, but SwarmUI’s native Qwen Image 2.1 support loads qwen3vl_8b_int8_convrot.safetensors; qwen3vl_8b.safetensors is the name SwarmUI uses for its Ideogram 4 encoder. The folder was right: the downloader puts text encoders in the folder SwarmUI’s own settings name (Models/text_encoders or Models/clip).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YqqbQCHG6vhdZTnE.png" /></figure><ul><li>V184 keeps the original names of all three Qwen Image 2.1 text encoders (INT8 ConvRot, BF16 and W4A8). The two Qwen Image 2.1 Unified presets and our QwenImage21 extension 1.0.2 select qwen3vl_8b_int8_convrot, so the presets and plain Qwen Image 2.1 generations share one 8.71 GB file. In our test the downloader got it in 1 min 28 s (1 min 46 s with the SHA-256 check), and SwarmUI then generated with it without downloading anything.</li></ul><p>Downloaded with an older version? Older downloaders saved every Qwen Image 2.1 text encoder, and Ideogram 4’s too, as qwen3vl_8b.safetensors. Check that file’s size in SwarmUI’s text encoder folder (Properties shows the bytes; a file SwarmUI has used is about 64 KB larger):</p><ul><li>9,350,798,360 bytes: the Qwen Image 2.1 INT8 ConvRot encoder. Rename it to qwen3vl_8b_int8_convrot.safetensors. If SwarmUI already downloaded that file itself, the old one is a spare 8.71 GB copy.</li><li>17,534,334,616 bytes: the Qwen Image 2.1 BF16 encoder, rename it to qwen3vl_8b_bf16.safetensors. 6,312,105,364 bytes: the W4A8 encoder, rename it to qwen3vl_8b_w4a8.safetensors.</li><li>10,588,637,512 or 6,305,221,764 bytes: Ideogram 4’s own FP8 or NVFP4 encoder: keep it.</li><li>After a rename, Ideogram 4 gets its own encoder from the Ideogram 4 Core Bundle or SwarmUI’s automatic download. The Qwen Image 2.1 Unified presets do not download models, so they need qwen3vl_8b_int8_convrot.safetensors: rename the old file, or run the Qwen Image 2.1 Core Bundle again (it keeps the model and the VAE after checking them).</li></ul><h3>UPDATED — REDESIGNED, MUCH FASTER MODEL DOWNLOADER</h3><p>V183 refreshed its status panels through a timer every half second, and after a search or Expand All every refresh blocked the browser for 0.2 to 0.35 seconds. V184 updates only what changed, and search and the precision filter no longer redraw the whole page. Measured in Chrome on our PC with the same catalog, V183 → V184:</p><ul><li>Search while typing: results shown about 0.44 seconds after the last key → already shown 0.12 seconds after it.</li><li>Clear Search: 673 ms → 10 ms. Expand All: 648 ms → 194 ms.</li><li>Precision filter “Display only GGUF models”: 851 ms → 11 ms.</li><li>Idle page after these steps: busy 65% of the time → 0%.</li><li>New look: bundle and model counts in the header, a light / dark switch that your browser remembers, model and bundle cards with an exact size on every file, colored buttons, a colored Activity Log and Active Transfers cards with a progress bar, size, speed and ETA (size and speed also while a file is being checked).</li><li>Search shows up to 50 result cards (was 15) with the category, the size or the bundle total; click a card to queue it.</li><li>Exact sizes: files below 0.1 GB show in MB, so a 4.8 MB file no longer reads “Size: 0 GB” and counts as unknown in the totals.</li><li>The status refresh pauses while the browser tab is hidden; downloads continue. The app’s download, queue and cancel logic is unchanged; the only change to file checks is the header check above.</li></ul><h3>NEW AND UPDATED PROMPT GUIDES FOR LLMS</h3><ul><li>New: Fibo_2.0_Enchance_Prompt_Feed_For_LLMs.txt. Paste it into ChatGPT, Claude, Gemini or a local LLM, describe the image you want or the edit (attach the images to the LLM chat for an edit), and it returns a JSON caption in Bria’s format. Its first part explains how the SwarmUI and ComfyUI presets use the prompt: where it goes, image labels, resolutions and the negative prompt. It is based on Bria’s model card and the caption schemas of Bria’s own captioning library, plus our tests.</li><li>MOSS_Image_Enchance_Prompt_Feed_For_LLMs.txt: new Part 7 for the MOSS Image Edit presets: one short plain-text instruction, the change first and then what stays, and how to name several images.</li><li>MiniMax_H3_Enchance_Prompt_Feed_For_LLMs.txt: new FastH3 Omni section: write FastH3 prompts exactly like MiniMax H3 references-to-video prompts.</li></ul><h3>UPDATED — BUNDLE REPORT AND FLUX 2 PRESET DESCRIPTIONS</h3><ul><li>Which_Bundles_Downloads_Which_Preset_Models.html is regenerated for V83: all 84 presets, and every new preset is fully covered by its core bundle. The report generator now knows FastH3 Omni’s automatic text encoder and VAEs, the edit LoRA that MOSS Image edit mode loads, Qwen Image 2.1’s own text encoder and VAE, and shows small files in MB.</li><li>The four FLUX 2 8-step presets told you to add — no-cache to the backend arguments for out-of-RAM errors. ComfyUI’s argument is — cache-none; their descriptions are corrected.</li></ul><h3>TESTED — REAL GENERATIONS IN SWARMUI</h3><p>Every result shown in this post is a real SwarmUI run on 23 October on an RTX 5090 (SwarmUI d74f8ce4, ComfyUI 6cbc3862) after Windows_Install_or_Update_SwarmUI.bat had installed the new extensions and nodes, with the V83 presets imported by Windows_Preset_Delete_Import.bat and applied by real clicks: all four Fibo 2.0 presets, both MOSS Image Edit presets and three FastH3 Omni presets, with seed 261023. The input portrait and four of the five MOSS seed checks ran earlier the same day, before that update. The FastH3 Omni Long Lip Sync preset was tested separately in SwarmUI with a portrait and 15 seconds of speech. The downloader ran from its own bat file: it kept the Fibo 2.0, MOSS-Image, Ideogram 4 and MiniMax H3 files that SwarmUI had rewritten, saved the Qwen Image 2.1 encoder under its own name, and the Qwen Image 2.1 Unified preset then generated without any download in SwarmUI.</p><h3>CUSTOM NODE PACKS WE DEVELOP OR MAINTAIN UPDATED</h3><p>SECoursesAudioTools</p><ul><li>New Fibo 2.0 nodes for our SwarmUI and ComfyUI Fibo 2.0 presets: prompt and image encoding with ComfyUI’s native Fibo 2.0 text encoder (JSON captions encoded like Bria’s pipeline, edits keep the first image’s size from 1 to 4 MP), Bria’s size-dependent shift, and masked edits with ComfyUI’s native Mask Editor.</li><li>ComfyUI now puts the Fibo 2.0 negative prompt into Bria’s chat template itself (22 October), so our nodes no longer wrap it a second time. With the same seed, the four Fibo 2.0 presets give pixel-identical images to before.</li><li>New MOSS-Image Edit References node: the reference branch of ChenkinLab’s Moss-Image-Edit adapter on ComfyUI’s native MOSS-Image model, used by the MOSS Image Edit presets in both apps.</li><li>H3 Long Lip Sync without a soundtrack now says that the soundtrack is needed instead of failing with a TypeError.</li></ul><p>ComfyUI-TeaCache</p><ul><li>The MiniMax H3 Speed Optimizer has a new VSA mode for FastH3 models: ComfyUI’s native Video Sparse Attention, used the way FastH3 was trained. Its output is bit-identical to ComfyUI’s own VSA node, and sampling is 1.37× (5 seconds) and 1.87× (10 seconds) faster than dense attention at 1344×768 on an RTX 5090.</li><li>The default face-pass prompt no longer asks for “beard strands”: on a FastH3 Omni image of a woman, the face pass had drawn a beard. Saved workflows keep their own text.</li></ul><h3>SWARMUI PREMIUM EXTENSIONS UPDATED</h3><ul><li>Fibo2 1.0.0 (new): Fibo 2 Advanced encoding like Bria’s pipeline, Bria’s size-dependent shift, automatic edit size (1 to 4 MP), masked edits from Init Image plus mask, the INT8 text encoder as default and prompt batches. It acts only on Fibo 2.0 models. Windows_Install_or_Update_SwarmUI.bat installs it.</li><li>MossImage: new Image edit (attached images) mode: it adds the Moss-Image-Edit LoRA and the reference branch, gives the result image 1’s reference size, turns @image1 into “image 1” and keeps the attached images out of SwarmUI’s text encoder.</li><li>MiniMaxH3References 1.21.0: the FastH3 sparse attention choice vsa; when SwarmUI’s NVFP4 Qwen3-VL 32B encoder is missing and the INT8 ConvRot encoder of the FastH3 Omni Core Bundle is installed, MiniMax H3 models use it instead of downloading another 14.61 GB file; Audio Only works with FastH3 Omni without a routing warning.</li><li>QwenImage21 1.0.2: the Qwen Unified Text Encoder default is qwen3vl_8b_int8_convrot, the file SwarmUI’s native Qwen Image 2.1 support loads.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v184.zip from this post and extract it over your existing downloader folder.</li><li>Run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI to the version with native Fibo 2.0 support (added on 21 October), installs the new Fibo2 premium extension, MossImage with Image edit mode, MiniMaxH3References 1.21.0 and QwenImage21 1.0.2, and updates SECoursesAudioTools and ComfyUI-TeaCache.</li><li>Update the ComfyUI backend to its latest version. Fibo 2.0 needs ComfyUI from 21 October or newer and FastH3 Omni from 18 October or newer. With the updated SECoursesAudioTools of step 2, Fibo 2 Dev text to image encodes its negative prompt exactly like Bria’s pipeline only on ComfyUI from late 22 October or newer. If SwarmUI uses our ComfyUI installation, update it with our ComfyUI installer V144 (Windows_Install_Or_Update_ComfyUI.bat). If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start. You can switch it back afterwards. Our ComfyUI installer V144 also brings the same Fibo 2.0, MOSS Image Edit and FastH3 Omni presets and the new prompt guides to ComfyUI.</li><li>Restart SwarmUI and its ComfyUI backend so the new extensions and nodes register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v83.json. It backs up your current presets, thumbnails included, to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the FastH3 Omni Core Bundle, the Fibo 2.0 Core Bundle and the MOSS-Image Core Bundle (the files you already have are checked and kept). Qwen Image 2.1 users: see the text encoder fix above.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>18 October 2026 Update V183</h3><p>This is a MOSS-Image update. V183 adds six MOSS-Image presets and our new MossImage premium extension to SwarmUI’s new native support for OpenMOSS’s MOSS-Image: the official quality recipe, Fast 25 Steps and Turbo 8 Steps presets, two Base model presets, visual layout boxes, JSON layout prompts and prompt batches from .txt files. MOSS-Image renders the text you quote in your prompt, such as shop signs, poster titles and labels. The model downloader gets the MOSS-Image Core Bundle (21.64 GB), the zip adds a MOSS-Image prompt-writing guide for LLMs, and AvatarForever gets live previews and new tested lip motion settings.</p><p>Amazing SwarmUI Presets moves to V81 with 74 presets: 6 new MOSS-Image presets, an updated AvatarForever preset and new artwork for the three FLUX 3 Klein presets; the other 64 presets are unchanged. The new presets need the new MOSS-Image Core Bundle (21.64 GB) of the model downloader.</p><p>All 74 presets as they appear in the SwarmUI Presets tab; the six MOSS-Image presets are in the last two rows:</p><p>The three FLUX 3 Klein presets just before them have new artwork as well (an otter, a flamingo and a meerkat), so every preset has its own animal again. The six new MOSS-Image presets on their own:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1c8f4oO-QHDfAMkD.png" /></figure><p>Like every other preset, each MOSS-Image preset has its own artwork that shows its task: Text To Image 50 Steps, Fast 25 Steps, Turbo 8 Steps, Layout Boxes, and the Base Model Text To Image and Layout Boxes presets.</p><h3>NEW PRESETS — MOSS-IMAGE: QUALITY, FAST, TURBO, LAYOUT BOXES AND BASE MODEL</h3><ul><li><em>“MOSS Image — Advanced Text To Image — 261017”</em></li><li><em>“MOSS Image — Fast 25 Steps — 261017”</em></li><li><em>“MOSS Image — Turbo 8 Steps — 261017”</em></li><li><em>“MOSS Image — Advanced Layout Boxes — 261017”</em></li><li><em>“MOSS Image Base — Advanced Text To Image — 261018”</em></li><li><em>“MOSS Image Base — Advanced Layout Boxes — 261018”</em></li></ul><p>All six load MOSS-Image in INT8 ConvRot with SwarmUI’s own Qwen3.5 9B INT8 ConvRot text encoder and the FLUX 2 VAE, set 1024×1024 and turn on our MossImage extension. MOSS-Image is a 5B single-stream image model that reads plain language, tags or JSON layouts, with prompts up to about 4K tokens. Measured in SwarmUI on an RTX 5090 at 1024×1024, for a new prompt with the models already loaded:</p><ul><li>Advanced Text To Image: the official quality recipe (50 Euler steps, normal scheduler, CFG 3, shift 5, negative-branch skip of layer 22 and the official negative tags) plus PyTorch attention and the SFT de-band fix. 12.9 seconds per image.</li><li>Fast 25 Steps: 25 ER-SDE steps with Score Rescaling 0.95 and layers 22 and 24 skipped from 10% to 70% of sampling: 6.9 seconds, nearly twice as fast. Large lettering held up in our tests; small text can get an occasional misspelling.</li><li>Turbo 8 Steps: the community 8-step DMD LoRA (lodestones, Apache-2.0) at 0.7 with 8 Euler steps and CFG 1: 1.8 seconds per image. The LoRA is a work in progress, so its images differ more from the official model and can show stray small text.</li><li>Advanced Layout Boxes: the official recipe with visual layout boxes, ready with three example boxes.</li><li>Base — Advanced Text To Image and Base — Advanced Layout Boxes: the raw Base model with the official 50-step recipe (13.0 seconds for Base Text To Image). Base gives more varied images and is the model that LoRAs trained on Base expect; SFT gives clearer text grouping and layouts. With the same seed, 80 steps gave no visible gain over 50 in our tests, so the Base presets keep 50. There is no Base Turbo preset, because the 8-step LoRA was trained on SFT.</li></ul><p>Where to find it: Presets tab → type MOSS Image in the filter → click a preset card → Apply. Its settings fill the new MOSS-Image group of the parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_LV8IhWAg-cMWfnN.png" /></figure><p>The MOSS-Image group and the preset settings in detail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xF4GQKtuffV4l5Cd.png" /></figure><ul><li>Seed: -1 gives a new image every time; a fixed seed repeats the same image (pixel-identical in our tests).</li><li>Mode: Text / JSON prompt or Layout boxes.</li><li>Negative Skip Layers (22): the official skip-layer guidance; the negative branch runs without these layers. Skip Start / End limit it to part of sampling. It needs CFG above 1.</li><li>Attention (PyTorch): with PyTorch attention the INT8 ConvRot files gave images close to the official BF16 outputs. Setting it here also protects the BF16 files, which gave black images with a — use-sage-attention backend in our RTX 3090 tests.</li><li>Score Rescaling (1 is off): the Fast preset uses 0.95; lower values give more detail, higher values smoother images.</li><li>SFT De-Band (on): MOSS-Image SFT draws faint vertical color stripes on skies, fog, walls and backdrops. The fix loads MATLOWAI’s de-band LoRA (Apache-2.0, in the MOSS-Image Core Bundle) for SFT models only; Base models never get it.</li><li>Your usual SwarmUI controls keep working: resolution, batch size, variation seed and LoRAs. 1 to 2 megapixel resolutions are recommended.</li><li>The extension acts only on MOSS-Image models. Apply any other preset and its own workflow runs unchanged.</li></ul><h3>NEW — ONE PROMPT, FOUR PRESETS: QUALITY, SPEED AND THE BASE MODEL</h3><p>The same prompt and seed (261018) through four of the new presets. The prompt asks for a bookshop on a rainy evening with a sign that reads “THE LAST CHAPTER”, a chalkboard that says “New arrivals / Hot tea inside”, a sleeping orange cat in the window and string lights:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yEJGtCoVqdAvvGmV.jpeg" /></figure><p>All four rendered the sign and the chalkboard text as written, with the sleeping cat in the window. Fast needs about half the time of the 50-step preset, and Turbo returns an image in 1.8 seconds, which suits trying many prompts. The 50-step presets are the most faithful; Base gives its own, more varied take on the same prompt.</p><h3>NEW — VISUAL LAYOUT BOXES: PUT TEXT AND OBJECTS WHERE YOU WANT THEM</h3><p>Choose Layout boxes mode, draw boxes on a canvas that has the shape of your output image and describe each box. The main prompt becomes the overall image caption. The extension builds the layout with ComfyUI’s native MOSS-Image layout nodes, so the model places each box’s content where you drew it.</p><ul><li>Draw a box with the mouse and drag it to move it; select a box and drag its lower-right corner to resize it.</li><li>Each box is an Object, Character or Text / lettering box. Text boxes render their exact text, line breaks included, and every edge can be typed in percent.</li><li>Add, delete and reorder up to 32 boxes, or open the JSON.</li><li>Boxes guide the composition; they are not inpainting masks.</li></ul><p>Where to find it: apply the MOSS Image — Advanced Layout Boxes preset or its Base version. The canvas is in the MOSS-Image group.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uDgdNCLdtGS7MpHm.png" /></figure><p>The preset’s three boxes put the title “EXPLORE THE ALPS” at the top, the mountain lake illustration in the middle and the subtitle “Summer 2026 | Lakes, trails and fresh air” at the bottom, as drawn, in 13.5 seconds of generation time.</p><h3>NEW — PROMPT SYNTAX AND JSON LAYOUT PROMPTS</h3><ul><li>MOSS-Image reads the prompt with Qwen3.5 as plain language, tags or JSON. Long, specific prompts work best; put the text that must appear in the image in quotes and use real line breaks to control its lines.</li><li>{red|blue} picks one at random, following the seed.</li><li>[red:blue:0.5] switches from one word to the other halfway through sampling, and [red|blue] alternates them from step to step.</li><li>(word:1.2) weights are removed, because Qwen3.5 would read them as literal text; describe emphasis in words instead. Times and ratios such as (10:30) or (4:3) stay.</li><li>A complete JSON layout passes through unchanged, and pixel drafts from the official MOSS-Image prompt enhancer are converted to the model’s own box format at your output size. SwarmUI’s own prompt tags keep working.</li></ul><h3>NEW — PROMPT BATCHES FROM .TXT FILES OR A WHOLE FOLDER</h3><p>Generate a series with the same settings. Add prompt adds a prompt you type; Import .txt and Import folder load text files, one complete prompt per file, JSON layouts included. Generate prompt batch runs your main prompt first and then every extra prompt; Stop after current ends the batch after the current prompt finishes. The normal Generate button still generates only the main prompt.</p><p>Where to find it: MOSS-Image group → MOSS Image Prompt List, in every MOSS-Image preset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DURe9xsTOS8t_jW9.png" /></figure><p>Three imported .txt files plus the main prompt gave four images with the Turbo 8 Steps preset in 7.4 seconds, from click to finish: the bookshop, a “COASTAL ESCAPE” travel poster, a “MORNING RITUAL” coffee cup and a sleeping fox. A batch takes up to 63 extra prompts. If a preset is only selected and not applied, the batch asks you to apply it first, so the preset’s own prompt cannot replace your batch prompts.</p><h3>NEW — MODEL DOWNLOADER: MOSS-IMAGE CORE BUNDLE (21.64 GB)</h3><p>One click downloads everything the six presets use:</p><ul><li>moss_image_sft_int8_convrot.safetensors (5.04 GB): the quality fine-tuned SFT model, the preset default.</li><li>moss_image_base_int8_convrot.safetensors (5.04 GB): the raw Base model for the Base presets, LoRA training and variety.</li><li>Qwen3.5 9B INT8 ConvRot text encoder (9.08 GB), saved as qwen3.5_9b_int8_convrot.safetensors, the file and name SwarmUI itself downloads for MOSS-Image, so SwarmUI reuses it.</li><li>FLUX 2 VAE (0.31 GB), saved as VAE/Flux/flux2-vae.safetensors, the same file FLUX 2 Klein and FLUX 3 Klein use.</li><li>MOSS_Image_SFT_DeBand_L37–41.safetensors (1.05 GB, MATLOWAI, Apache-2.0): the SFT banding fix that the presets apply automatically.</li><li>MOSS_Image_SFT_8_Steps_DMD_LoRA.safetensors (1.11 GB, lodestones, Apache-2.0): the 8-step LoRA of the Turbo preset, saved from lodestones’ latest release file, which can change upstream.</li></ul><p>Where to find it: run Windows_Start_Download_Models_App.bat. The MOSS-Image Core Bundle is the first bundle of the SwarmUI Bundles tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yu7riGKCEhO7nd0a.png" /></figure><p>Optional, under Image Generation Models → MOSS-Image Models: the original BF16 SFT and Base models (9.47 GB each; BF16 Base is the recommended precision for LoRA training and fine-tuning), a smaller Qwen3.5 9B W4A8 GPTQ text encoder for low RAM and VRAM (6.84 GB; select it in the Qwen Model parameter), the FLUX 2 small-decoder VAE that decodes faster with less VRAM (0.23 GB; select it as the VAE) and an extra banding fix for stubborn smooth skies (1.05 GB).</p><p>Why INT8 ConvRot with PyTorch attention: on an RTX 3090 the INT8 ConvRot files run about twice as fast as BF16, and with PyTorch attention their images were close to the official BF16 outputs. The INT8 files’ own INT8 attention is about 20% faster, but it changed the images noticeably in our tests, so the presets use PyTorch attention. Our ComfyUI graph with these settings reproduced official MOSS-Image Space examples from the same seed, with the same composition, subject and details.</p><h3>UPDATED — AVATARFOREVER: LIVE PREVIEWS, TESTED LIP MOTION SETTINGS AND A HISTORY FIX</h3><ul><li>Live previews: the AvatarForever sampler now shows a preview frame after every denoising step, starting at step 1, and the finished video after export; in SwarmUI the previews appear in the normal generation preview. They use taeltx2_3.safetensors (0.02 GB) when you have it, now part of the LTX 2.3 Avatar Forever Core Bundle (44.53 GB), and a fast latent preview otherwise. Previews do not change the video: with previews on and off, the same 452-frame MP4 came out byte for byte in our test.</li><li>Tested lip motion settings: the AvatarForever Unified Image + Audio or Text + Audio preset now renders an exact 512×768 portrait canvas with 16 latent frames per chunk (was 4) and a prompt that asks for expressive mouth movement. In a six-second comparison with the same portrait, song, prompt and seed, 4-frame chunks kept the lips closed and 16-frame chunks gave visible mouth movement, which held through the complete 18-second song. At 1024×1024 the mouth stayed mostly closed in the same test even with 16-frame chunks, so the preset keeps 512×768. Larger chunks use more memory; check a short sample when you change the image, the audio or the resolution.</li><li>History fix (AvatarForever 1.0.3): an interrupted video export could leave an incomplete MP4 that kept SwarmUI’s History waiting, and with several browser tabs open even finished images stopped loading. SwarmUI now skips the metadata of such files; complete files and generation settings are unchanged.</li></ul><p>Where to find it: Presets tab → type AvatarForever → click the preset card → Apply, then attach a speech or song to the prompt and, if you like, a portrait as Init Image.</p><h3>NEW — MOSS-IMAGE PROMPT GUIDE FOR LLMS</h3><p>New file: Automatic_Prompt_Generate_Enhance_To_Video_And_Image_Models_LLM_Instructions/MOSS_Image_Enchance_Prompt_Feed_For_LLMs.txt. Give it to ChatGPT, Claude, Gemini or a local LLM together with your idea, and it writes MOSS-Image prompts: detailed captions, exact text for signs and posters, and layout drafts with boxes that the presets convert to MOSS-Image’s own coordinates at your output size. Its first part explains how the SwarmUI and ComfyUI presets use the prompt (recipe, prompt syntax, layout mode and resolutions); it is based on the official MOSS-Image user guide and its prompt enhancer.</p><h3>UPDATED — BUNDLE REPORT AND DOWNLOADER APP</h3><ul><li>Which_Bundles_Downloads_Which_Preset_Models.html is regenerated for V81: all 74 presets, and every MOSS-Image preset is fully covered by the MOSS-Image Core Bundle. The report generator now knows SwarmUI’s default MOSS-Image text encoder and VAE, the de-band LoRA that the MossImage extension loads by itself for SFT models, and AvatarForever’s live-preview file.</li><li>Windows_Start_Download_Models_App.bat now installs the downloader’s packages in pip’s isolated mode, so your user pip configuration and PIP_ option variables no longer affect its install.</li><li>The downloader app shows v183.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v183.zip from this post and extract it over your existing downloader folder.</li><li>Run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI to the version with native MOSS-Image support (added on 17 October), installs the new MossImage premium extension and AvatarForever 1.0.3, and updates SECoursesAudioTools.</li><li>Update the ComfyUI backend: MOSS-Image needs ComfyUI from 16 October or newer. If SwarmUI uses our ComfyUI installation, update it with our ComfyUI installer V143 (Windows_Install_Or_Update_ComfyUI.bat). If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start. You can switch it back afterwards. Our ComfyUI installer V143 also brings the same six MOSS-Image presets and the MOSS-Image prompt guide to ComfyUI.</li><li>Restart SwarmUI and its ComfyUI backend so the new extension and nodes register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v81.json. It backs up your current presets, thumbnails included, to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the MOSS-Image Core Bundle. AvatarForever users: also download taeltx2_3.safetensors (0.02 GB) for the live previews; it is in the LTX 2.3 Avatar Forever Core Bundle and under Video Generation Models → LTX 2.3 Video Models.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>16 October 2026 Update V182</h3><p>This is a FLUX 3 Klein update. V182 brings Black Forest Labs’ new FLUX 3 Klein image model to SwarmUI with three advanced presets and our new FLUX 3 Klein premium extension: text to image, image editing with up to 10 reference images, visual layout boxes, ComfyUI-style prompt syntax and prompt batches from .txt files. The model downloader gets the FLUX 3 Klein Core Bundle with our HQ INT8 ConvRot model (10.85 GB in total), and the zip adds a FLUX 3 Klein prompt-writing guide for LLMs.</p><p>Amazing SwarmUI Presets moves to V80 with 68 presets: 3 new FLUX 3 Klein presets, and the other 65 presets are unchanged. The new presets need the new FLUX 3 Klein Core Bundle (10.85 GB) of the model downloader.</p><p>All 68 presets as they appear in the SwarmUI Presets tab; the three FLUX 3 Klein presets are in the last row:</p><p>Like every other preset, each FLUX 3 Klein preset has its own artwork that shows its task: Text To Image, Image Edit or Layout Boxes.</p><h3>NEW PRESETS — FLUX 3 KLEIN: TEXT TO IMAGE, IMAGE EDIT AND LAYOUT BOXES</h3><p><em>“FLUX 3 Klein — Advanced Text To Image — 261015”</em></p><p><em>“FLUX 3 Klein — Advanced Image Edit — 261015”</em></p><p><em>“FLUX 3 Klein — Advanced Layout Boxes — 261015”</em></p><p>All three load our HQ INT8 ConvRot FLUX 3 Klein model with SwarmUI’s own Qwen3-VL 4B text encoder and the FLUX 2 VAE, and set 1024×1024, 16 steps, Euler/simple and CFG 1 (CFG 1 ignores the negative prompt). Each preset turns on our FLUX 3 Klein extension in its own mode: Text / optional references, Reference edit or Layout / optional references. Measured in SwarmUI on an RTX 5090 at 1024×1024: about 2 seconds of generation time for an image and 3.9 seconds for an edit with one reference image.</p><p>Where to find it: Presets tab → type FLUX 3 Klein in the filter → click a preset card → Apply. Its settings fill the new FLUX.3 Klein group of the parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8OAo-QzEbHYsgmCv.png" /></figure><p>The FLUX.3 Klein group and the preset settings in detail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*S1KINs9gyaTkh-7N.png" /></figure><ul><li>Seed: -1 gives a new image every time; a fixed seed repeats the same image (pixel-identical in our tests).</li><li>Mode: text to image, reference edit or layout boxes. Reference images are optional in the text and layout modes.</li><li>Reference Megapixels: the size of each attached image, from 0.1 to 4 megapixels (1 is about 1024×1024), independent of the output size.</li><li>Your usual SwarmUI controls keep working: resolution, batch size, variation seed and compatible FLUX 3 Klein LoRAs.</li><li>The extension acts only on FLUX 3 Klein models. Apply any other preset and its own workflow runs unchanged.</li></ul><h3>NEW — EDIT IMAGES WITH UP TO 10 REFERENCE IMAGES (@image1, @image2)</h3><p>Attach images to the prompt with Add Image / Video / Audio and cite them in your prompt: @image1 is the first attachment, @image2 the second, and so on (@img1, @picture1 and @pic1 work too). Describe the complete result: what changes and what stays. The extension turns these names into FLUX 3 Klein’s own reference labels.</p><ul><li>Up to 10 images. Each keeps its aspect ratio, and transparent areas are filled with white.</li><li>Reference edit needs at least one image. If the prompt cites an image that is not attached, SwarmUI shows a clear error; video and audio attachments are refused with a message.</li></ul><p>Where to find it: apply the FLUX 3 Klein — Advanced Image Edit preset, attach your photo above the prompt and describe the result.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8gDXNIssqOHIAWFq.png" /></figure><p>Here the white mug of the attached photo became cobalt blue in 3.9 seconds of generation time, while the croissant, the window, the light and the camera angle stayed. To take one object out of several references, name the image it comes from and say which other objects should be absent: in our test with this coffee photo as @image1 and a red teapot as @image2, asking for the vessel from @image2 gave the teapot alone.</p><h3>NEW — VISUAL LAYOUT BOXES: PUT EVERY OBJECT WHERE YOU WANT IT</h3><p>Draw boxes on a canvas that has the shape of your output image, describe the object of each box, and cite the boxes in your prompt as &lt;obj_1&gt;, &lt;obj_2&gt;. SwarmUI turns them into FLUX 3 Klein’s native layout table, so the model places each object in its box.</p><ul><li>Draw a box with the mouse and drag it to move it; select a box and drag its lower-right corner to resize it.</li><li>Each box is an Object or Text / lettering with the exact text to render, and its edges can be typed in percent.</li><li>Add, delete and reorder up to 32 boxes, or open the JSON. Reference images still work in this mode.</li><li>Boxes guide the composition; they are not inpainting masks.</li></ul><p>Where to find it: apply the FLUX 3 Klein — Advanced Layout Boxes preset. The canvas is in the FLUX.3 Klein group.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*q6mTKUCMuZf6LYPT.png" /></figure><p>The preset’s two boxes put the red teapot on the left and the blue cup on the right, as drawn, in 2.0 seconds of generation time.</p><h3>NEW — COMFYUI-STYLE PROMPT SYNTAX</h3><p>The FLUX 3 Klein presets understand the prompt syntax of our ComfyUI presets, next to SwarmUI’s own prompt tags:</p><ul><li>(ceramic:1.1) weights a phrase.</li><li>[red:blue:0.5] switches from one word to the other halfway through sampling.</li><li>[red|blue] alternates the words from step to step.</li><li>&lt;break&gt; splits the prompt into separately encoded parts.</li><li>{red|blue} picks one at random, following the seed.</li><li>SwarmUI tags such as &lt;random:red|blue&gt;, &lt;fromto[0.5]:red|blue&gt; and &lt;alternate:red|blue&gt; keep working.</li><li>A complete native caption with its JSON layout table, as the new prompt guide writes it, works in the text and reference edit modes.</li></ul><h3>NEW — PROMPT BATCHES FROM .TXT FILES OR A WHOLE FOLDER</h3><p>Generate a series with the same settings and reference images. Add prompt adds a prompt you type; Import .txt and Import folder load text files, one complete prompt per file, in natural file-name order. Generate prompt batch runs your main prompt first and then every extra prompt; Stop after current ends the batch after the current prompt finishes. The normal Generate button still generates only the main prompt.</p><p>Where to find it: FLUX.3 Klein group → FLUX Klein Prompt List, in every FLUX 3 Klein preset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uvQVdHSeRnADFvx_.png" /></figure><p>Three imported .txt files plus the main prompt gave four images in 8 seconds, from click to finish. A batch takes up to 63 extra prompts.</p><h3>NEW — MODEL DOWNLOADER: FLUX 3 KLEIN CORE BUNDLE (10.85 GB)</h3><p>One click downloads everything the three presets use:</p><ul><li>FLUX-3-Image-Klein-INT8-ConvRot-HQ.safetensors (5.66 GB): our HQ INT8 ConvRot version of the distilled FLUX 3 Klein, with selected layers kept in BF16.</li><li>Qwen3-VL 4B FP8 text encoder (4.88 GB), saved as qwen3vl_4b.safetensors, the file name SwarmUI itself uses for FLUX 3 Klein.</li><li>FLUX 2 VAE (0.31 GB), saved as VAE/Flux/flux2-vae.safetensors, the same file FLUX 2 Klein uses: if you already have it, it is verified and skipped.</li></ul><p>Where to find it: run Windows_Start_Download_Models_App.bat. The FLUX 3 Klein Core Bundle is the first bundle of the SwarmUI Bundles tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HzwA3UXQXLD3I1ad.png" /></figure><p>Optional, under Image Generation Models → FLUX 3 Klein Models: the original distilled BF16 model (10.34 GB), the undistilled Base BF16 checkpoint for fine-tuning and LoRA training (10.32 GB; accept its license on Black Forest Labs’ gated Hugging Face page and enter your Hugging Face token in the downloader; it uses its own 50-step, guidance 3.5 settings), and a Qwen3-VL 4B BF16 text encoder (8.27 GB). The BF16 encoder has the same qwen3vl_4b.safetensors name as the FP8 one, so download only one of them.</p><p>Why HQ INT8: in our 40-state test against the original BF16 model (BF16 text encoder, RTX 3090), the denoiser output of our HQ INT8 ConvRot model differed by 2.14% on average, against 3.05% for the official INT8 ConvRot file with its default settings and 2.27% with matched attention; with matched attention, one held-out edit case favored the official file (3.84% against 3.58%; 4.73% with its default settings). The transformer peaked at about 6 GiB of VRAM instead of 10.5 GiB and ran each step about twice as fast as BF16. These are numerical-fidelity measurements, not image-quality percentages.</p><h3>UPDATED — TEXT ENCODERS GO WHERE YOUR SWARMUI LOOKS FOR THEM</h3><p>SwarmUI keeps text encoders in the folders of its SDClipFolder setting: text_encoders;clip on installs from late November 2025 on, and only clip on older installs, because SwarmUI never changes a setting it has saved. The downloader now reads this setting from your SwarmUI (Data/Settings.fds next to the Models folder you download into):</p><ul><li>A text encoder you already have in one of those folders is verified where it is, instead of being downloaded a second time.</li><li>A new text encoder goes into the first folder, where SwarmUI saves its own downloads too.</li><li>If the downloader cannot find your SwarmUI settings, it uses clip, which SwarmUI scans by default on every install.</li></ul><p>The ComfyUI and Forge folders are unchanged. We tested this on new, old and custom SwarmUI folder settings.</p><h3>NEW — FLUX 3 KLEIN PROMPT GUIDE FOR LLMS</h3><p>New file: Automatic_Prompt_Generate_Enhance_To_Video_And_Image_Models_LLM_Instructions/FLUX_3_Klein_Enchance_Prompt_Feed_For_LLMs.txt. Give it to ChatGPT, Claude, Gemini or a local LLM together with your idea, and it writes FLUX 3 Klein prompts: text-to-image captions, edits that say what stays and what changes, multi-reference prompts with @image1 and @image2, exact text for posters and signs, and layout boxes. Its first part explains the SwarmUI presets (modes, @image numbering, the box editor, prompt syntax and batches); the rest is the same research-based guide, version 1.5, as in our ComfyUI installer, with original examples.</p><h3>UPDATED — BUNDLE REPORT, PRESET BACKUPS AND DOWNLOADER APP</h3><ul><li>Which_Bundles_Downloads_Which_Preset_Models.html is regenerated for V80: all 68 presets, and the three FLUX 3 Klein presets are fully covered by the FLUX 3 Klein Core Bundle. The report generator now also knows SwarmUI’s default FLUX 3 Klein text encoder and VAE.</li><li>Preset backups keep their thumbnails: Windows_Preset_Delete_Import.bat now saves every preset’s thumbnail inside the backup file, so a backup holds everything needed to bring your own presets back with their images.</li><li>The downloader app shows v182 and uses Gradio 6.30.0. Windows_Start_Download_Models_App.bat updates Gradio automatically, and the RunPod / SimplePod and Massed Compute instructions install it.</li></ul><h3>TESTED — REAL GENERATIONS IN CHROME</h3><p>The extension and the presets were tested on 15 October with SwarmUI 967f005c and ComfyUI 20b46065 on an RTX 5090 in a real Chrome: text to image, scene-preserving edits, two-reference selection, visual layout (drawing, moving and resizing boxes), advanced prompt syntax, prompt batches, .txt file and folder import, Stop after current, and 23 workflow checks (reference numbering, the 10-image limit, duplicate images, layout tables, reference sizing and model gating). Every generated result shown in this post is a real run on the HQ INT8 model. The optional BF16 models and FLUX 3 Klein LoRAs were not generated separately.</p><h3>CUSTOM NODE PACKS WE DEVELOP OR MAINTAIN UPDATED</h3><p>SECoursesAudioTools</p><ul><li>LoRA and DoRA files trained with SECourses Musubi Trainer now load completely. ComfyUI silently skipped some of their layers: the MLP halves of Qwen Image 2.1, the q, k and v parts of Z-Image’s fused attention, and every HiDream-O1 layer. The node pack now maps the trainer’s layer names to them, DoRA scales included. SwarmUI gets this too: Windows_Install_or_Update_SwarmUI.bat installs this node pack for SwarmUI’s self-starting ComfyUI backend.</li><li>New FLUX.3 Gallery Prompt and FLUX.3 Optional Gallery References nodes connect our reference gallery to FLUX 3 Klein in our ComfyUI presets: @image1 names, seeded {red|blue} choices and up to 10 references with their own size.</li></ul><h3>SWARMUI PREMIUM EXTENSIONS UPDATED</h3><ul><li>FluxThreeKlein 1.0.0 (new): FLUX 3 Klein advanced prompting, image references, visual layout boxes and prompt batches on SwarmUI’s native FLUX 3 workflow. Windows_Install_or_Update_SwarmUI.bat installs it.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v182.zip from this post and extract it over your existing downloader folder.</li><li>Run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI to the version with native FLUX 3 Klein support (added on 15 October), installs the new FluxThreeKlein premium extension and updates SECoursesAudioTools.</li><li>Update the ComfyUI backend: FLUX 3 Klein needs ComfyUI from 15 October or newer (tested with 20b46065). If SwarmUI uses our ComfyUI installation, update it with our ComfyUI installer V141 (Windows_Install_Or_Update_ComfyUI.bat). If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start. You can switch it back afterwards.</li><li>Restart SwarmUI and its ComfyUI backend so the new extension and nodes register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v80.json. It backs up your current presets, thumbnails included, to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the FLUX 3 Klein Core Bundle.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fef73fc4d600" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Higgsfield Genjutsu (Character Replace) is Now Available With SwarmUI v185 and ComfyUI v145 locally…]]></title>
            <link>https://medium.com/@furkangozukara/higgsfield-genjutsu-character-replace-is-now-available-with-swarmui-v185-and-comfyui-v145-locally-db8c40e8e658?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/db8c40e8e658</guid>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 04:12:00 GMT</pubDate>
            <atom:updated>2026-10-24T04:12:36.225Z</atom:updated>
            <content:encoded><![CDATA[<h3>Higgsfield Genjutsu (Character Replace) is Now Available With SwarmUI v185 and ComfyUI v145 locally for free via MiniMax H3–1 click to install and use</h3><h3>Demo video here : <a href="https://www.patreon.com/SECourses/posts/higgsfield-is-h3-173243286">https://www.patreon.com/SECourses/posts/higgsfield-is-h3-173243286</a></h3><h3>ComfyUI Download : <a href="https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709">https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709</a></h3><h3>SwarmUI Download : <a href="https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862">https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862</a></h3><ul><li>Use MiniMax H3 Core Bundle from SwarmUI Model Downloader to have all models</li></ul><h3>ComfyUI V145 ChangeLogs</h3><h3>24 October 2026 Update V145</h3><p>Presets Folder Content In Latest Zip File</p><p>This is a MiniMax H3 Genjutsu update. V145 adds two Genjutsu presets that replace the person in any video with your own character: add your clip and one image of the new character, press Run, and the result keeps the clip’s motion, timing, camera, length, aspect ratio and original sound. One preset uses FastH3 Omni at 8 steps, the other MiniMax H3 Ref2VA at 20 steps. V145 also adds a Genjutsu section to the MiniMax H3 prompt guide and brings the new Genjutsu nodes of our FoleyExtension.</p><p>Our test: this armoured warrior image and a 9.8-second phone clip of a man shadowboxing in a gym. Both new presets put the warrior into the clip, punch for punch and step for step, in the clip’s portrait format and with its own sound:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BhvfqeMsORuL4z8N.jpeg" /></figure><p>Every result in this post comes from a real run of the shipped presets in ComfyUI started with our Windows_Run_ComfyUI_Default.bat: each preset was opened from the Workflows sidebar and run with the Run button, with its shipped settings and the example it opens with. The times are ComfyUI’s times on our RTX 5090 for a first run, model loading included. We ran both presets again on the newest ComfyUI that the installer installs today and got the same videos, frame for frame.</p><p>Exact V144 → V145 zip audit:</p><ul><li>2 new files: the two MiniMax H3 Genjutsu presets.</li><li>1 changed file: the MiniMax H3 prompt guide, with the new Genjutsu section.</li><li>No files were removed. The other 238 files are identical. The Genjutsu nodes arrive through Windows_Install_Or_Update_ComfyUI.bat, which installs our updated FoleyExtension.</li></ul><p>Here is the complete Presets folder of V145: 108 ready-to-use ComfyUI presets, plus 3 API presets in the api_paid folder. The two new presets are marked NEW:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*k7u5up1WUOIY4JjC.png" /></figure><p>The presets cover image, video, lip sync, audio and music models: AvatarForever, ERNIE Image, FastH3 Omni, Fibo 2.0, FLUX.1 Dev, Kontext and SRPO, FLUX 2, FLUX 3 Klein, Ideogram 4, Krea 2, LTX 2, LTX 2.3, LTX 2.5, MiniMax H3 (now with Genjutsu), MiniMax Music 3, MOSS-Image and MOSS Image Edit, Phantom Wan, Qwen Image, SDXL, Stable Video Infinity 2 Pro, Wan 2.1, Wan 2.2 and Z-Image.</p><h3>NEW PRESETS — MINIMAX H3 GENJUTSU: PUT YOUR CHARACTER INTO ANY VIDEO</h3><p>“MiniMax H3 Genjutsu — FastH3 Omni 8 Steps — 261024”<br>“MiniMax H3 Genjutsu — Ref2VA 20 Steps — 261024”</p><p>How it works: SAM 3.1 finds the person in every frame of your clip, and only that person is colour-inverted. MiniMax H3 then copies the clip’s motion, timing, camera and lip movements, but not the performer’s face, hair or clothes. A local Qwen3-VL 8B (Auto Describe) looks at your character image and six frames of the clip and writes the prompt for you: the new outfit, the action, the setting, the lighting and the framing. The clip’s soundtrack is locked into MiniMax H3’s audio stream for lip sync, and the saved video carries the original sound.</p><ul><li>Your inputs: the clip as @video1 and one image of the new character as @image1, added with Add references or Load + trim (to pick the part of the clip you want; 4 to 10 seconds works best, up to 15). The presets open with our example loaded; if those example files are not on your PC, remove the two cards and add your own.</li><li>The output follows the clip: the same length, the same aspect ratio at about 1 megapixel, and the original sound. Our 9.8-second 1080x1920 clip gave 226 frames (9.4 seconds) at 768x1344.</li><li>Prompt: keep the short default note, type your own notes (for example “He wears his red scarf.”), empty it, or paste a complete MiniMax H3 prompt from ChatGPT, Claude or Gemini; a complete prompt is used as written.</li><li>Background: video keeps the clip’s location, lighting and framing and replaces only the performer (character replacement); image keeps your character image’s background and takes only the motion (motion transfer).</li><li>Target: who to replace, in plain words: person, or man in red:1 to pick one of several people.</li><li>Speed: FastH3 Omni 8 steps made the 9.4-second example in 4.4 minutes and Ref2VA 20 steps in 8.3 to 9.4 minutes, first runs with model loading on an RTX 5090.</li></ul><p>Here is the FastH3 Omni preset right after the run of our example, with every part numbered:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BbagdtSUpVAtgK7z.png" /></figure><p>After a run, CHECK THE MASK shows the performer in red in every frame, MOTION GUIDE shows what MiniMax H3 copies the motion from, and FINAL PROMPT SENT TO H3 shows the prompt Auto Describe wrote. The token meter under the prompt now follows Genjutsu too: it shows the clip’s length and aspect ratio before you press Run.</p><h3>UPDATED PROMPT GUIDE — MINIMAX H3 GENJUTSU</h3><p>The MiniMax H3 guide in Automatic_Prompt_Generate_Enhance_To_Video_And_Image_Models_LLM_Instructions has a new Genjutsu section. Give it to an LLM that can watch video (for example Gemini) together with your clip and your character image, and paste its answer into the prompt box. It can write the exact spoken words for word-by-word lip sync.</p><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V145.zip from this post.</li><li>Extract it over your existing installer folder and overwrite the older files.</li><li>Close ComfyUI, and SwarmUI too if it uses this ComfyUI as its backend, then run Windows_Install_Or_Update_ComfyUI.bat. It updates ComfyUI and installs our FoleyExtension with the Genjutsu nodes.</li><li>If you have not done it since V144, run Windows_Custom_Nodes_Bundles_Installer.bat and install/update bundle 100.</li><li>Restart ComfyUI and refresh the browser tab.</li></ol><p>Use our latest SwarmUI Model Downloader (SwarmUI_Model_Downloader_v185.zip or newer) and download the bundle from the SwarmUI Bundles tab:</p><ul><li>FastH3 Omni Core Bundle (60.12 GB, 11 files) for the FastH3 Omni Genjutsu preset, or the MiniMax-H3 Core Bundle (79.91 GB, 19 files) for the Ref2VA 20 steps preset. Both now include SAM 3.1 (1.63 GB) and Qwen3-VL 8B INT8 (8.71 GB, the same file as the Qwen Image 2.1 encoder, so you do not download it twice), sizes as the downloader shows them.</li><li>If your ComfyUI reads your SwarmUI Models folder through extra_model_paths.yaml, ComfyUI uses the files downloaded there without a second download.</li></ul><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>SwarmUI V185 Logs</h3><h3>24 October 2026 Update V185</h3><p>This is a MiniMax H3 Genjutsu update. V185 adds two presets that replace the person in any video with your own character: attach your clip and one image of the new character, press Generate, and the result keeps the clip’s motion, timing, camera, length, aspect ratio and original sound. Our MiniMaxH3References extension gets the new MiniMax H3 Genjutsu group, and the model downloader adds SAM 3.1 and Qwen3-VL 8B to the MiniMax H3 and FastH3 Omni bundles.</p><p>Amazing SwarmUI Presets moves to V84 with 86 presets: the two new presets, and the other 84 presets unchanged.</p><p>Our test in SwarmUI: this armoured warrior image and a 9.8-second phone clip of a man shadowboxing in a gym. Both presets put the warrior into the clip, punch for punch and step for step, in the clip’s portrait format and with its own sound:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9NTRP3D33-2bp4iK.jpeg" /></figure><p>Exact V184 → V185 zip audit:</p><ul><li>Presets: Amazing_SwarmUI_Presets_v83.json was replaced by Amazing_SwarmUI_Presets_v84.json. Its 84 presets are unchanged in every setting, description and thumbnail; 2 presets are new.</li><li>6 changed files: the model catalog and its size data (SAM 3.1 and Qwen3-VL 8B in the MiniMax-H3 Core, MiniMax-H3 Low VRAM and FastH3 Omni Core bundles), update_swarmui.py (a premium extension that could not be replaced during an update is installed again by the next update), the downloader app (shows v185), the regenerated Which_Bundles_Downloads_Which_Preset_Models.html and the MiniMax H3 prompt guide (new Genjutsu section).</li></ul><h3>NEW PRESETS — MINIMAX H3 GENJUTSU: PUT YOUR CHARACTER INTO ANY VIDEO</h3><p>“MiniMax H3 Genjutsu — FastH3 Omni 8 Steps — 261024”<br>“MiniMax H3 Genjutsu — Ref2VA 20 Steps — 261024”</p><p>How it works: SAM 3.1 finds the person in every frame of your clip, and only that person is colour-inverted. MiniMax H3 then copies the clip’s motion, timing, camera and lip movements, but not the performer’s face, hair or clothes. A local Qwen3-VL 8B (Auto Describe) looks at your character image and six frames of the clip and writes the prompt for you: the new outfit, the action, the setting, the lighting and the framing. The clip’s soundtrack is locked into MiniMax H3’s audio stream for lip sync, and the saved video carries the original sound.</p><p>Where to find it: Quick Tools → Reset Params to Default, then the Presets tab → type Genjutsu in the filter → the preset card’s menu → Direct Apply. Attach the clip and the character image with Add References (or Add A Reference With Trim to pick the part of the clip) and press Generate. Here is the FastH3 Omni preset after a real run of our example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GWht0Sp-5CQEwNAR.png" /></figure><ul><li>MiniMax H3 Genjutsu (on): turns the first reference video into the SAM 3.1 motion guide.</li><li>Genjutsu Background: video (character replacement) keeps the clip’s location, lighting and framing; image (motion transfer) keeps your character image’s background and takes only the motion.</li><li>Genjutsu Target: who to replace, in plain words: person, or man in red:1 to pick one of several people.</li><li>Genjutsu Auto Describe (on): the local Qwen3-VL 8B writes the outfit, action, setting, lighting and framing into the prompt. Leave the prompt empty, type short notes, or paste a complete MiniMax H3 prompt from ChatGPT, Claude or Gemini; a complete prompt is used as written.</li><li>Genjutsu Spoken Words (optional): the words spoken in the clip, for word-by-word lip sync.</li><li>Genjutsu Lock Clip Audio (on) and Genjutsu Match Clip Aspect (on): the clip’s sound drives the lip movements, and the output takes the clip’s aspect ratio at your Width x Height area.</li><li>Genjutsu Motion Guide Megapixels (0.15): the guide size; advanced options set the detection threshold and the mask grow.</li><li>The output follows the clip: our 9.8-second 1080x1920 clip gave 226 frames (9.4 seconds) at 768x1344, and the token meter beside the prompt shows that length and size before you press Generate.</li><li>Speed: FastH3 Omni 8 steps made the 9.4-second example in 4.5 minutes and Ref2VA 20 steps in 9.9 minutes in SwarmUI, first runs with model loading on an RTX 5090.</li></ul><h3>MODEL DOWNLOADER — SAM 3.1 AND QWEN3-VL 8B IN THE MINIMAX H3 BUNDLES</h3><p>The MiniMax-H3 Core Bundle, the MiniMax-H3 Low VRAM Bundle and the FastH3 Omni Core Bundle now download everything the Genjutsu presets use:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xhrNSCVgHfYIHNa4.png" /></figure><ul><li>SAM 3.1 Multiplex FP16 (1.63 GB, saved to Models/Stable-Diffusion): finds the person to replace in every frame.</li><li>Qwen3-VL 8B INT8 ConvRot (8.71 GB, saved to Models/text_encoders): writes the Genjutsu prompt. It is the same file as the Qwen Image 2.1 encoder, so if you already have it, the downloader checks it and keeps it.</li><li>New totals: MiniMax-H3 Core Bundle 79.91 GB (19 files), MiniMax-H3 Low VRAM Bundle 64.21 GB (19 files), FastH3 Omni Core Bundle 60.12 GB (11 files), sizes as the downloader shows them.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v185.zip from this post and extract it over your existing downloader folder.</li><li>Close SwarmUI and run Windows_Install_or_Update_SwarmUI.bat. It installs MiniMaxH3References 1.22.1 and our FoleyExtension with the Genjutsu nodes.</li><li>Keep the ComfyUI backend up to date. If SwarmUI uses our ComfyUI installation, close SwarmUI again (step 2 starts it, and a running backend blocks the update) and update that ComfyUI with our ComfyUI installer V145 (Windows_Install_Or_Update_ComfyUI.bat), which also brings the same Genjutsu presets to ComfyUI. If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start.</li><li>Restart SwarmUI and its ComfyUI backend so the new extension and nodes register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v84.json. It backs up your current presets, thumbnails included, to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the MiniMax-H3 Core Bundle or the FastH3 Omni Core Bundle (the files you already have are checked and kept).</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db8c40e8e658" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Latest SwarmUI Updates — Amazing New Models and Presets]]></title>
            <link>https://medium.com/@furkangozukara/latest-swarmui-updates-amazing-new-models-and-presets-847390a8fab0?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/847390a8fab0</guid>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 14:21:45 GMT</pubDate>
            <atom:updated>2026-10-23T14:21:45.238Z</atom:updated>
            <content:encoded><![CDATA[<h3>Download The Zip File From Here : <a href="https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862">https://www.patreon.com/SECourses/posts/swarmui-auto-and-114517862</a></h3><h3>Here below latest 4 updates we made</h3><h3>23 October 2026 Update V184</h3><p>This is a Fibo 2.0, MOSS Image Edit and FastH3 Omni update. V184 adds four presets for Bria’s Fibo 2.0, which generates and edits images from long JSON captions, two MOSS Image Edit presets that edit your photos with a short instruction, and four FastH3 Omni presets that turn your images, videos and sounds into video with synchronized audio, stills or audio in 8 steps. The model downloader gets the FastH3 Omni and Fibo 2.0 Core Bundles and a redesigned, much faster interface, keeps the model files whose header SwarmUI rewrote instead of downloading them again, and fixes the Qwen Image 2.1 text encoder that SwarmUI downloaded a second time.</p><p>Amazing SwarmUI Presets moves to V83 with 84 presets: 10 new presets, corrected descriptions for the four FLUX 2 8-step presets and the fixed text encoder in the two Qwen Image 2.1 Unified presets; the other 68 presets are unchanged. The new presets need the new FastH3 Omni Core Bundle, the Fibo 2.0 Core Bundle and the updated MOSS-Image Core Bundle of the model downloader.</p><p>All 84 presets as they appear in the SwarmUI Presets tab; the new presets are in the last rows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rVLqGuhZUxr-6tAS.jpeg" /></figure><p>Every preset still has its own animal. The ten new presets on their own:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mT7GuaEl-ngbznRW.png" /></figure><p>Four Fibo 2.0 presets (text to image and image editing, Dev and Turbo), two MOSS Image Edit presets (30 steps and Turbo 8 steps) and four FastH3 Omni presets (references to video, long lip sync, text to image and audio only), each with its own artwork.</p><p>Exact V183 → V184 zip audit:</p><ul><li>Presets: Amazing_SwarmUI_Presets_v81.json was replaced by Amazing_SwarmUI_Presets_v83.json. Of the 74 presets it keeps, 68 are unchanged in every setting, description and thumbnail; the four FLUX 2 8-step presets have a corrected description and the two Qwen Image 2.1 Unified presets select the fixed text encoder.</li><li>9 changed files: the model catalog and its size data (the FastH3 Omni and Fibo 2.0 models and bundles, the MOSS-Image edit LoRA, the Qwen Image 2.1 encoder names and the new source of the MiniMax H3 video VAE), the downloader app (new interface, shows v184), HF_model_downloader.py (keeps files whose header SwarmUI rewrote), update_swarmui.py (the new Fibo2 premium extension), the bundle report generator and the regenerated Which_Bundles_Downloads_Which_Preset_Models.html, and the MOSS-Image and MiniMax H3 prompt guides.</li><li>New: the Fibo 2.0 prompt guide in Automatic_Prompt_Generate_Enhance_To_Video_And_Image_Models_LLM_Instructions.</li><li>Windows_Install_or_Update_SwarmUI.bat brings SwarmUI with its native Fibo 2.0 support, the new Fibo2 premium extension, MossImage with Image edit mode, MiniMaxH3References 1.21.0, QwenImage21 1.0.2, SECoursesAudioTools and ComfyUI-TeaCache. The ComfyUI backend is updated separately (step 3 of HOW TO UPDATE).</li></ul><h3>NEW PRESETS — FIBO 2.0: TEXT TO IMAGE AND IMAGE EDITING, DEV AND TURBO</h3><p><em>“Fibo 2 Dev — Advanced Text To Image — 261022”</em></p><p><em>“Fibo 2 Dev — Advanced Image Edit — 261022”</em></p><p><em>“Fibo 2 Turbo — Image Edit 4 Steps — 261022”</em></p><p><em>“Fibo 2 Turbo — Text To Image 4 Steps — 261022”</em></p><p>Fibo 2.0 by Bria is an 8B image model that generates and edits images. It was trained on long structured JSON captions that describe every object, the lighting, the camera and the text in the image, so a JSON caption gives the most control; plain text also works. All four presets load the models in INT8 ConvRot with the Qwen3-VL 4B INT8 ConvRot text encoder and the FLUX 2 VAE, set 1024×1024 and turn on our new Fibo2 extension. Measured in SwarmUI on an RTX 5090 at 1024×1024, for a new prompt with the models already loaded:</p><ul><li>Dev — Advanced Text To Image: Bria’s official recipe: 50 Euler steps, CFG 5 and Bria’s JSON negative prompt from the model card. 17.3 seconds per image.</li><li>Turbo — Text To Image 4 Steps: Bria made the 4-step Edit Turbo model for editing, but in our tests it also renders text to image very well, lettering included: 4 steps, CFG 1, 1.9 seconds per image.</li><li>Dev — Advanced Image Edit: the official editing recipe with 50 steps and CFG 5: 34.7 seconds.</li><li>Turbo — Image Edit 4 Steps: Bria’s distilled editing model, 4 steps, CFG 1: 2.5 seconds.</li></ul><p>About 1 megapixel is the trained size (1024×1024, 1424×800, 1248×832 and the other shapes); 2 megapixels such as 1664×1248 also work well.</p><p>Where to find it: Presets tab → type Fibo 2 in the filter → click a preset card → Apply. Its settings fill the new Fibo 2.0 group of the parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Rfzu8WD7hOH3Bz7h.png" /></figure><ul><li>Fibo 2 Advanced (on): encodes your prompt with ComfyUI’s native Fibo 2.0 text encoder exactly like Bria’s pipeline. SwarmUI’s generic encoder wraps Fibo 2.0 prompts differently and leaves edit images out of the text encoder.</li><li>Fibo 2 Sigma Shift: Bria’s official shift, which changes with the image size.</li><li>Fibo 2 Edit Size: edits keep the size of the first image from 1 to 4 megapixels.</li><li>Editing: attach 1 to 4 images to the prompt with Add Image / Video / Audio (or drag them onto the prompt box) and write the instruction; @image1, @image2 refer to them. Masked edit: an Init Image plus a mask painted in the image editor (white = change).</li><li>Prompt syntax: a JSON caption in Bria’s format (the new Fibo 2.0 LLM guide writes them for you) or plain text. {red|blue} picks one at random following the seed; (word:1.2) weights are removed because Qwen3-VL would read them as text.</li><li>Fibo 2 Prompt List: prompt batches from typed prompts, .txt files or a whole folder (one prompt per file, JSON captions included), like the MOSS-Image prompt batches. Generate prompt batch runs your main prompt first, then every extra prompt.</li><li>The extension acts only on Fibo 2.0 models. Apply any other preset and its own workflow runs unchanged.</li></ul><h3>NEW — ONE JSON PROMPT: FIBO 2 DEV 50 STEPS VS TURBO 4 STEPS</h3><p>The same JSON caption and seed (261023) through both text-to-image presets. The caption asks for a cafe on a rainy evening with a green sign that reads “MORNING BREW” and a sidewalk chalkboard with “Fresh Coffee” and “Open Late”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Xn5wfFx_DhjV-IuC.jpeg" /></figure><p>Both presets wrote every word as asked, with the warm window light, the wet cobblestones and the guests inside. Turbo needs 1.9 seconds instead of 17.3, which suits trying many prompts; Dev is the full, non-distilled model with CFG and Bria’s official recipe, and the model to train LoRAs on.</p><h3>NEW — ONE PORTRAIT, THREE NEW EDIT PRESETS</h3><p>A portrait made with Fibo 2 Turbo Text To Image, attached to the prompt as Image 1 and edited with each new edit preset and its default instruction (seed 261023):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tAEPRVV5kMt55Aul.jpeg" /></figure><p>Both Fibo 2.0 edits redrew the photo as an anime illustration and kept the pose, the clothes, the coffee cup and the flower shop. MOSS Image Edit replaced the street with a sunny beach and kept the same face, hair, clothes and pose, in 11.1 seconds.</p><h3>NEW PRESETS — MOSS IMAGE EDIT: EDIT YOUR PHOTOS WITH ONE SENTENCE</h3><p><em>“MOSS Image Edit — Advanced Image Edit — 261022”</em></p><p><em>“MOSS Image Edit — Turbo 8 Steps — 261022”</em></p><p>Both presets use MOSS-Image SFT INT8 ConvRot with ChenkinLab’s Moss-Image-Edit LoRA, which the MossImage extension’s new Image edit mode adds by itself together with the adapter’s reference branch. Attach up to 4 images: image 1 is edited and the others are references, named image 1, image 2 (or @image1, @image2) in the instruction. Say what changes and what stays, for example “Change the background to a sunny beach, preserving the person’s identity and pose.”</p><ul><li>Advanced Image Edit: the official edit recipe: 30 steps, CFG 4, shift 5 and an empty negative prompt. 11.1 seconds at 1024×1024 on an RTX 5090.</li><li>Turbo 8 Steps: the community 8-step DMD LoRA at 0.7 with 8 steps and CFG 1: 2.4 seconds. Turbo is close to the 30-step preset in many edits, but for our busy street portrait it kept the street on all five seeds we tried, while the 30-step preset gave a beach on all five. When Turbo ignores a big change such as a new background, use the 30-step preset.</li><li>The result has image 1’s shape at up to 1024 pixels on the long side (MOSS Image Edit Reference Size; 1536 also works).</li><li>Colors, clothes, expressions, lighting, seasons, backgrounds and style edit best. Placing an object from a second image works; copying its exact look is weaker.</li></ul><p>Where to find it: Presets tab → type MOSS Image Edit → click a preset card → Apply, then add the photo with Add Image / Video / Audio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eooGhoAEyciRpmrt.png" /></figure><p>The preset sets MOSS Image Mode to Image edit (attached images). The new MOSS_Image_Edit_LoRA.safetensors (0.28 GB) is part of the updated MOSS-Image Core Bundle.</p><h3>NEW PRESETS — FASTH3 OMNI: VIDEO WITH AUDIO FROM YOUR REFERENCES IN 8 STEPS</h3><p><em>“FastH3 Omni — References to Video 8 Steps — 261022”</em></p><p><em>“FastH3 Omni — Long Lip Sync — References 8 Steps — 261022”</em></p><p><em>“FastH3 Omni — Text to Image with Optional References — 261022”</em></p><p><em>“FastH3 Omni — Audio Only with Optional References — 261022”</em></p><p>FastH3 Omni by FastVideo and Nuva Lab is MiniMax H3’s references-to-video-and-audio model distilled to 8 steps. One model makes video with synchronized 32 kHz stereo audio from your prompt plus up to 9 images, 3 videos and 3 audio files, with no turbo LoRA: 8 steps, CFG 1. The presets load it in INT8 ConvRot with the INT8 ConvRot Qwen3-VL 32B text and vision encoder. The References to Video and Long Lip Sync presets also run the sparse attention FastH3 was trained with (VSA): sampling was about 1.4× faster for 5 seconds and 1.9× for 10 seconds of 1344×768 video on an RTX 5090, with the same quality in our tests.</p><ul><li>References to Video: 1344×768, 124 frames (5.2 seconds) with audio: 53 to 57 seconds in our runs with one reference image (one run took 108 seconds; we could not find why). It also works as text to video without references.</li><li>Long Lip Sync: put the complete soundtrack in Init Audio and a portrait beside the prompt as @image1; one 24 FPS video follows the whole audio in overlapping 8-step windows at 832×1248. About 3.5 minutes for 15 seconds of speech (3.32 minutes in our SwarmUI test).</li><li>Text to Image with Optional References: stills from text, or with up to 9 reference images: 9.8 seconds at 1344×768 for a new prompt with one reference image.</li><li>Audio Only with Optional References: 32 kHz stereo audio returned as a lossless FLAC: 8 seconds of audio in 7.2 seconds. Put spoken words inside &lt;d&gt;[English] … &lt;/d&gt;.</li></ul><p>Where to find it: first Quick Tools → Reset Params to Default, so no MiniMax H3 turbo LoRA from an earlier preset stays selected. Then Presets tab → type FastH3 Omni → click a preset card → Apply, add your images, videos or audio with Add References and mention them in the prompt as @image1, @video1, @audio1.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7N6VNktM_6a-NQsH.png" /></figure><p>The prompt describes the woman from @image1 in front of a flower shop: she smiles, raises her coffee cup, greets the viewer with a spoken line in &lt;d&gt;, laughs and takes a sip. Nine frames from the start to the end of the clip:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*te2S3HEhpRVMaq6k.jpeg" /></figure><p>Her face, curly hair, freckles, earrings and clothes stay those of the portrait in every frame, and the clip follows the prompt from the greeting to the sip. Video and its 32 kHz stereo soundtrack came out of one 8-step run.</p><h3>NEW — MODEL DOWNLOADER: FASTH3 OMNI AND FIBO 2.0 CORE BUNDLES</h3><ul><li>FastH3 Omni Core Bundle (49.78 GB, 9 files): FastH3 Omni INT8 ConvRot (20.69 GB), the MiniMax H3 Qwen3-VL 32B INT8 ConvRot text and vision encoder (25.28 GB), the MiniMax H3 video VAE (2.62 GB) and audio VAE (0.56 GB), plus the face models the presets’ face refinement and mouth pass use. MiniMax H3 users already have the audio VAE and the face models, and the INT8 ConvRot encoder if they chose it. The video VAE now comes from Comfy-Org, the file SwarmUI itself downloads: if an earlier downloader saved Kijai’s copy at the same path, it is replaced once (2.62 GB).</li><li>Fibo 2.0 Core Bundle (20.50 GB, 4 files): Fibo 2.0 Dev and Edit Turbo in INT8 ConvRot (7.83 GB each), the Qwen3-VL 4B INT8 ConvRot text encoder (4.52 GB, saved as qwen3vl_4b_int8_convrot.safetensors so it never replaces the qwen3vl_4b.safetensors other models use) and the shared FLUX 2 VAE (0.31 GB).</li><li>MOSS-Image Core Bundle (21.91 GB, 7 files): now also downloads MOSS_Image_Edit_LoRA.safetensors (0.28 GB) for the MOSS Image Edit presets.</li></ul><p>Where to find it: run Windows_Start_Download_Models_App.bat. The FastH3 Omni Core Bundle and the Fibo 2.0 Core Bundle are the first two bundles of the SwarmUI Bundles tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RW5tdhauW3j99kru.png" /></figure><p>Optional, under Video Generation Models → FastH3 Omni Video Models: FastH3 Omni in BF16 (41.20 GB), FP8 (21.46 GB) and NVFP4 QAD (12.83 GB, for RTX 50 series and other Blackwell GPUs). Under Image Generation Models → Fibo 2.0 Models: both models in BF16 (15.32 GB each), the BF16 text encoder (8.27 GB, saved as qwen3vl_4b.safetensors, the name the FP8 encoder of FLUX 3 Klein and Krea 2 also uses, so keep one precision) and Comfy-Org’s Fibo 2.0 turbo LoRA (1.85 GB), which lets Dev and your Dev LoRAs run in 4 steps; the presets do not select it.</p><h3>NEW — THE DOWNLOADER KEEPS MODELS WHOSE HEADER SWARMUI REWROTE</h3><p>The first time SwarmUI uses a model, it writes the model hash into the file’s header (SwarmUI’s ImageMetadataIncludeModelHash setting, on by default). The tensor data stays the same, but the file gets a little larger or its bytes change, so V183 failed its size or SHA-256 check and downloaded the whole model again. V184 reads the original header from Hugging Face, checks your file’s tensor data behind it, and keeps the file when every tensor byte matches. A checked file is remembered, so the next run skips it at once.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TlJ344ICM1cpHvSL.png" /></figure><p>The log says “[VERIFIED] … the tensor data matches the Hugging Face file; only the header metadata differs” and “[SKIP] … already complete”. On our test PC, 59 model files with 581 GB in total were a little larger than their Hugging Face files, the sign of SwarmUI’s header rewrite. The Fibo 2.0 Core Bundle’s four files were checked and kept in 25 seconds instead of downloading 20.50 GB again. Files whose tensor data differs are still downloaded again, as before.</p><h3>FIXED — SWARMUI NO LONGER DOWNLOADS THE QWEN IMAGE 2.1 TEXT ENCODER AGAIN</h3><p>Thank you to the supporter who reported this: after the Qwen Image 2.1 Core Bundle, SwarmUI said the text encoder was missing and downloaded it again. The downloader saved the INT8 ConvRot Qwen3-VL 8B encoder as qwen3vl_8b.safetensors, but SwarmUI’s native Qwen Image 2.1 support loads qwen3vl_8b_int8_convrot.safetensors; qwen3vl_8b.safetensors is the name SwarmUI uses for its Ideogram 4 encoder. The folder was right: the downloader puts text encoders in the folder SwarmUI’s own settings name (Models/text_encoders or Models/clip).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0I-wCsPErPdfrGG4.png" /></figure><ul><li>V184 keeps the original names of all three Qwen Image 2.1 text encoders (INT8 ConvRot, BF16 and W4A8). The two Qwen Image 2.1 Unified presets and our QwenImage21 extension 1.0.2 select qwen3vl_8b_int8_convrot, so the presets and plain Qwen Image 2.1 generations share one 8.71 GB file. In our test the downloader got it in 1 min 28 s (1 min 46 s with the SHA-256 check), and SwarmUI then generated with it without downloading anything.</li></ul><p>Downloaded with an older version? Older downloaders saved every Qwen Image 2.1 text encoder, and Ideogram 4’s too, as qwen3vl_8b.safetensors. Check that file’s size in SwarmUI’s text encoder folder (Properties shows the bytes; a file SwarmUI has used is about 64 KB larger):</p><ul><li>9,350,798,360 bytes: the Qwen Image 2.1 INT8 ConvRot encoder. Rename it to qwen3vl_8b_int8_convrot.safetensors. If SwarmUI already downloaded that file itself, the old one is a spare 8.71 GB copy.</li><li>17,534,334,616 bytes: the Qwen Image 2.1 BF16 encoder, rename it to qwen3vl_8b_bf16.safetensors. 6,312,105,364 bytes: the W4A8 encoder, rename it to qwen3vl_8b_w4a8.safetensors.</li><li>10,588,637,512 or 6,305,221,764 bytes: Ideogram 4’s own FP8 or NVFP4 encoder: keep it.</li><li>After a rename, Ideogram 4 gets its own encoder from the Ideogram 4 Core Bundle or SwarmUI’s automatic download. The Qwen Image 2.1 Unified presets do not download models, so they need qwen3vl_8b_int8_convrot.safetensors: rename the old file, or run the Qwen Image 2.1 Core Bundle again (it keeps the model and the VAE after checking them).</li></ul><h3>UPDATED — REDESIGNED, MUCH FASTER MODEL DOWNLOADER</h3><p>V183 refreshed its status panels through a timer every half second, and after a search or Expand All every refresh blocked the browser for 0.2 to 0.35 seconds. V184 updates only what changed, and search and the precision filter no longer redraw the whole page. Measured in Chrome on our PC with the same catalog, V183 → V184:</p><ul><li>Search while typing: results shown about 0.44 seconds after the last key → already shown 0.12 seconds after it.</li><li>Clear Search: 673 ms → 10 ms. Expand All: 648 ms → 194 ms.</li><li>Precision filter “Display only GGUF models”: 851 ms → 11 ms.</li><li>Idle page after these steps: busy 65% of the time → 0%.</li><li>New look: bundle and model counts in the header, a light / dark switch that your browser remembers, model and bundle cards with an exact size on every file, colored buttons, a colored Activity Log and Active Transfers cards with a progress bar, size, speed and ETA (size and speed also while a file is being checked).</li><li>Search shows up to 50 result cards (was 15) with the category, the size or the bundle total; click a card to queue it.</li><li>Exact sizes: files below 0.1 GB show in MB, so a 4.8 MB file no longer reads “Size: 0 GB” and counts as unknown in the totals.</li><li>The status refresh pauses while the browser tab is hidden; downloads continue. The app’s download, queue and cancel logic is unchanged; the only change to file checks is the header check above.</li></ul><h3>NEW AND UPDATED PROMPT GUIDES FOR LLMS</h3><ul><li>New: Fibo_2.0_Enchance_Prompt_Feed_For_LLMs.txt. Paste it into ChatGPT, Claude, Gemini or a local LLM, describe the image you want or the edit (attach the images to the LLM chat for an edit), and it returns a JSON caption in Bria’s format. Its first part explains how the SwarmUI and ComfyUI presets use the prompt: where it goes, image labels, resolutions and the negative prompt. It is based on Bria’s model card and the caption schemas of Bria’s own captioning library, plus our tests.</li><li>MOSS_Image_Enchance_Prompt_Feed_For_LLMs.txt: new Part 7 for the MOSS Image Edit presets: one short plain-text instruction, the change first and then what stays, and how to name several images.</li><li>MiniMax_H3_Enchance_Prompt_Feed_For_LLMs.txt: new FastH3 Omni section: write FastH3 prompts exactly like MiniMax H3 references-to-video prompts.</li></ul><h3>UPDATED — BUNDLE REPORT AND FLUX 2 PRESET DESCRIPTIONS</h3><ul><li>Which_Bundles_Downloads_Which_Preset_Models.html is regenerated for V83: all 84 presets, and every new preset is fully covered by its core bundle. The report generator now knows FastH3 Omni’s automatic text encoder and VAEs, the edit LoRA that MOSS Image edit mode loads, Qwen Image 2.1’s own text encoder and VAE, and shows small files in MB.</li><li>The four FLUX 2 8-step presets told you to add — no-cache to the backend arguments for out-of-RAM errors. ComfyUI’s argument is — cache-none; their descriptions are corrected.</li></ul><h3>TESTED — REAL GENERATIONS IN SWARMUI</h3><p>Every result shown in this post is a real SwarmUI run on 23 October on an RTX 5090 (SwarmUI d74f8ce4, ComfyUI 6cbc3862) after Windows_Install_or_Update_SwarmUI.bat had installed the new extensions and nodes, with the V83 presets imported by Windows_Preset_Delete_Import.bat and applied by real clicks: all four Fibo 2.0 presets, both MOSS Image Edit presets and three FastH3 Omni presets, with seed 261023. The input portrait and four of the five MOSS seed checks ran earlier the same day, before that update. The FastH3 Omni Long Lip Sync preset was tested separately in SwarmUI with a portrait and 15 seconds of speech. The downloader ran from its own bat file: it kept the Fibo 2.0, MOSS-Image, Ideogram 4 and MiniMax H3 files that SwarmUI had rewritten, saved the Qwen Image 2.1 encoder under its own name, and the Qwen Image 2.1 Unified preset then generated without any download in SwarmUI.</p><h3>CUSTOM NODE PACKS WE DEVELOP OR MAINTAIN UPDATED</h3><p>SECoursesAudioTools</p><ul><li>New Fibo 2.0 nodes for our SwarmUI and ComfyUI Fibo 2.0 presets: prompt and image encoding with ComfyUI’s native Fibo 2.0 text encoder (JSON captions encoded like Bria’s pipeline, edits keep the first image’s size from 1 to 4 MP), Bria’s size-dependent shift, and masked edits with ComfyUI’s native Mask Editor.</li><li>ComfyUI now puts the Fibo 2.0 negative prompt into Bria’s chat template itself (22 October), so our nodes no longer wrap it a second time. With the same seed, the four Fibo 2.0 presets give pixel-identical images to before.</li><li>New MOSS-Image Edit References node: the reference branch of ChenkinLab’s Moss-Image-Edit adapter on ComfyUI’s native MOSS-Image model, used by the MOSS Image Edit presets in both apps.</li><li>H3 Long Lip Sync without a soundtrack now says that the soundtrack is needed instead of failing with a TypeError.</li></ul><p>ComfyUI-TeaCache</p><ul><li>The MiniMax H3 Speed Optimizer has a new VSA mode for FastH3 models: ComfyUI’s native Video Sparse Attention, used the way FastH3 was trained. Its output is bit-identical to ComfyUI’s own VSA node, and sampling is 1.37× (5 seconds) and 1.87× (10 seconds) faster than dense attention at 1344×768 on an RTX 5090.</li><li>The default face-pass prompt no longer asks for “beard strands”: on a FastH3 Omni image of a woman, the face pass had drawn a beard. Saved workflows keep their own text.</li></ul><h3>SWARMUI PREMIUM EXTENSIONS UPDATED</h3><ul><li>Fibo2 1.0.0 (new): Fibo 2 Advanced encoding like Bria’s pipeline, Bria’s size-dependent shift, automatic edit size (1 to 4 MP), masked edits from Init Image plus mask, the INT8 text encoder as default and prompt batches. It acts only on Fibo 2.0 models. Windows_Install_or_Update_SwarmUI.bat installs it.</li><li>MossImage: new Image edit (attached images) mode: it adds the Moss-Image-Edit LoRA and the reference branch, gives the result image 1’s reference size, turns @image1 into “image 1” and keeps the attached images out of SwarmUI’s text encoder.</li><li>MiniMaxH3References 1.21.0: the FastH3 sparse attention choice vsa; when SwarmUI’s NVFP4 Qwen3-VL 32B encoder is missing and the INT8 ConvRot encoder of the FastH3 Omni Core Bundle is installed, MiniMax H3 models use it instead of downloading another 14.61 GB file; Audio Only works with FastH3 Omni without a routing warning.</li><li>QwenImage21 1.0.2: the Qwen Unified Text Encoder default is qwen3vl_8b_int8_convrot, the file SwarmUI’s native Qwen Image 2.1 support loads.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v184.zip from this post and extract it over your existing downloader folder.</li><li>Run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI to the version with native Fibo 2.0 support (added on 21 October), installs the new Fibo2 premium extension, MossImage with Image edit mode, MiniMaxH3References 1.21.0 and QwenImage21 1.0.2, and updates SECoursesAudioTools and ComfyUI-TeaCache.</li><li>Update the ComfyUI backend to its latest version. Fibo 2.0 needs ComfyUI from 21 October or newer and FastH3 Omni from 18 October or newer. With the updated SECoursesAudioTools of step 2, Fibo 2 Dev text to image encodes its negative prompt exactly like Bria’s pipeline only on ComfyUI from late 22 October or newer. If SwarmUI uses our ComfyUI installation, update it with our ComfyUI installer V144 (Windows_Install_Or_Update_ComfyUI.bat). If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start. You can switch it back afterwards. Our ComfyUI installer V144 also brings the same Fibo 2.0, MOSS Image Edit and FastH3 Omni presets and the new prompt guides to ComfyUI.</li><li>Restart SwarmUI and its ComfyUI backend so the new extensions and nodes register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v83.json. It backs up your current presets, thumbnails included, to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the FastH3 Omni Core Bundle, the Fibo 2.0 Core Bundle and the MOSS-Image Core Bundle (the files you already have are checked and kept). Qwen Image 2.1 users: see the text encoder fix above.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>18 October 2026 Update V183</h3><p>This is a MOSS-Image update. V183 adds six MOSS-Image presets and our new MossImage premium extension to SwarmUI’s new native support for OpenMOSS’s MOSS-Image: the official quality recipe, Fast 25 Steps and Turbo 8 Steps presets, two Base model presets, visual layout boxes, JSON layout prompts and prompt batches from .txt files. MOSS-Image renders the text you quote in your prompt, such as shop signs, poster titles and labels. The model downloader gets the MOSS-Image Core Bundle (21.64 GB), the zip adds a MOSS-Image prompt-writing guide for LLMs, and AvatarForever gets live previews and new tested lip motion settings.</p><p>Amazing SwarmUI Presets moves to V81 with 74 presets: 6 new MOSS-Image presets, an updated AvatarForever preset and new artwork for the three FLUX 3 Klein presets; the other 64 presets are unchanged. The new presets need the new MOSS-Image Core Bundle (21.64 GB) of the model downloader.</p><p>All 74 presets as they appear in the SwarmUI Presets tab; the six MOSS-Image presets are in the last two rows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*66C2aJ-2J9bZ-NEY.jpeg" /></figure><p>The three FLUX 3 Klein presets just before them have new artwork as well (an otter, a flamingo and a meerkat), so every preset has its own animal again. The six new MOSS-Image presets on their own:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ppws1dtqbQeclukH.png" /></figure><p>Like every other preset, each MOSS-Image preset has its own artwork that shows its task: Text To Image 50 Steps, Fast 25 Steps, Turbo 8 Steps, Layout Boxes, and the Base Model Text To Image and Layout Boxes presets.</p><h3>NEW PRESETS — MOSS-IMAGE: QUALITY, FAST, TURBO, LAYOUT BOXES AND BASE MODEL</h3><ul><li><em>“MOSS Image — Advanced Text To Image — 261017”</em></li><li><em>“MOSS Image — Fast 25 Steps — 261017”</em></li><li><em>“MOSS Image — Turbo 8 Steps — 261017”</em></li><li><em>“MOSS Image — Advanced Layout Boxes — 261017”</em></li><li><em>“MOSS Image Base — Advanced Text To Image — 261018”</em></li><li><em>“MOSS Image Base — Advanced Layout Boxes — 261018”</em></li></ul><p>All six load MOSS-Image in INT8 ConvRot with SwarmUI’s own Qwen3.5 9B INT8 ConvRot text encoder and the FLUX 2 VAE, set 1024×1024 and turn on our MossImage extension. MOSS-Image is a 5B single-stream image model that reads plain language, tags or JSON layouts, with prompts up to about 4K tokens. Measured in SwarmUI on an RTX 5090 at 1024×1024, for a new prompt with the models already loaded:</p><ul><li>Advanced Text To Image: the official quality recipe (50 Euler steps, normal scheduler, CFG 3, shift 5, negative-branch skip of layer 22 and the official negative tags) plus PyTorch attention and the SFT de-band fix. 12.9 seconds per image.</li><li>Fast 25 Steps: 25 ER-SDE steps with Score Rescaling 0.95 and layers 22 and 24 skipped from 10% to 70% of sampling: 6.9 seconds, nearly twice as fast. Large lettering held up in our tests; small text can get an occasional misspelling.</li><li>Turbo 8 Steps: the community 8-step DMD LoRA (lodestones, Apache-2.0) at 0.7 with 8 Euler steps and CFG 1: 1.8 seconds per image. The LoRA is a work in progress, so its images differ more from the official model and can show stray small text.</li><li>Advanced Layout Boxes: the official recipe with visual layout boxes, ready with three example boxes.</li><li>Base — Advanced Text To Image and Base — Advanced Layout Boxes: the raw Base model with the official 50-step recipe (13.0 seconds for Base Text To Image). Base gives more varied images and is the model that LoRAs trained on Base expect; SFT gives clearer text grouping and layouts. With the same seed, 80 steps gave no visible gain over 50 in our tests, so the Base presets keep 50. There is no Base Turbo preset, because the 8-step LoRA was trained on SFT.</li></ul><p>Where to find it: Presets tab → type MOSS Image in the filter → click a preset card → Apply. Its settings fill the new MOSS-Image group of the parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mcQx5Sy6QFPP90fv.png" /></figure><p>The MOSS-Image group and the preset settings in detail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*osCOkFmOQi4RgMAE.png" /></figure><ul><li>Seed: -1 gives a new image every time; a fixed seed repeats the same image (pixel-identical in our tests).</li><li>Mode: Text / JSON prompt or Layout boxes.</li><li>Negative Skip Layers (22): the official skip-layer guidance; the negative branch runs without these layers. Skip Start / End limit it to part of sampling. It needs CFG above 1.</li><li>Attention (PyTorch): with PyTorch attention the INT8 ConvRot files gave images close to the official BF16 outputs. Setting it here also protects the BF16 files, which gave black images with a — use-sage-attention backend in our RTX 3090 tests.</li><li>Score Rescaling (1 is off): the Fast preset uses 0.95; lower values give more detail, higher values smoother images.</li><li>SFT De-Band (on): MOSS-Image SFT draws faint vertical color stripes on skies, fog, walls and backdrops. The fix loads MATLOWAI’s de-band LoRA (Apache-2.0, in the MOSS-Image Core Bundle) for SFT models only; Base models never get it.</li><li>Your usual SwarmUI controls keep working: resolution, batch size, variation seed and LoRAs. 1 to 2 megapixel resolutions are recommended.</li><li>The extension acts only on MOSS-Image models. Apply any other preset and its own workflow runs unchanged.</li></ul><h3>NEW — ONE PROMPT, FOUR PRESETS: QUALITY, SPEED AND THE BASE MODEL</h3><p>The same prompt and seed (261018) through four of the new presets. The prompt asks for a bookshop on a rainy evening with a sign that reads “THE LAST CHAPTER”, a chalkboard that says “New arrivals / Hot tea inside”, a sleeping orange cat in the window and string lights:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OkV9045SQB7AHqzr.jpeg" /></figure><p>All four rendered the sign and the chalkboard text as written, with the sleeping cat in the window. Fast needs about half the time of the 50-step preset, and Turbo returns an image in 1.8 seconds, which suits trying many prompts. The 50-step presets are the most faithful; Base gives its own, more varied take on the same prompt.</p><h3>NEW — VISUAL LAYOUT BOXES: PUT TEXT AND OBJECTS WHERE YOU WANT THEM</h3><p>Choose Layout boxes mode, draw boxes on a canvas that has the shape of your output image and describe each box. The main prompt becomes the overall image caption. The extension builds the layout with ComfyUI’s native MOSS-Image layout nodes, so the model places each box’s content where you drew it.</p><ul><li>Draw a box with the mouse and drag it to move it; select a box and drag its lower-right corner to resize it.</li><li>Each box is an Object, Character or Text / lettering box. Text boxes render their exact text, line breaks included, and every edge can be typed in percent.</li><li>Add, delete and reorder up to 32 boxes, or open the JSON.</li><li>Boxes guide the composition; they are not inpainting masks.</li></ul><p>Where to find it: apply the MOSS Image — Advanced Layout Boxes preset or its Base version. The canvas is in the MOSS-Image group.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bBpg8-GKUX3Q-dmP.png" /></figure><p>The preset’s three boxes put the title “EXPLORE THE ALPS” at the top, the mountain lake illustration in the middle and the subtitle “Summer 2026 | Lakes, trails and fresh air” at the bottom, as drawn, in 13.5 seconds of generation time.</p><h3>NEW — PROMPT SYNTAX AND JSON LAYOUT PROMPTS</h3><ul><li>MOSS-Image reads the prompt with Qwen3.5 as plain language, tags or JSON. Long, specific prompts work best; put the text that must appear in the image in quotes and use real line breaks to control its lines.</li><li>{red|blue} picks one at random, following the seed.</li><li>[red:blue:0.5] switches from one word to the other halfway through sampling, and [red|blue] alternates them from step to step.</li><li>(word:1.2) weights are removed, because Qwen3.5 would read them as literal text; describe emphasis in words instead. Times and ratios such as (10:30) or (4:3) stay.</li><li>A complete JSON layout passes through unchanged, and pixel drafts from the official MOSS-Image prompt enhancer are converted to the model’s own box format at your output size. SwarmUI’s own prompt tags keep working.</li></ul><h3>NEW — PROMPT BATCHES FROM .TXT FILES OR A WHOLE FOLDER</h3><p>Generate a series with the same settings. Add prompt adds a prompt you type; Import .txt and Import folder load text files, one complete prompt per file, JSON layouts included. Generate prompt batch runs your main prompt first and then every extra prompt; Stop after current ends the batch after the current prompt finishes. The normal Generate button still generates only the main prompt.</p><p>Where to find it: MOSS-Image group → MOSS Image Prompt List, in every MOSS-Image preset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qxwz8EV44Q_6rPI8.png" /></figure><p>Three imported .txt files plus the main prompt gave four images with the Turbo 8 Steps preset in 7.4 seconds, from click to finish: the bookshop, a “COASTAL ESCAPE” travel poster, a “MORNING RITUAL” coffee cup and a sleeping fox. A batch takes up to 63 extra prompts. If a preset is only selected and not applied, the batch asks you to apply it first, so the preset’s own prompt cannot replace your batch prompts.</p><h3>NEW — MODEL DOWNLOADER: MOSS-IMAGE CORE BUNDLE (21.64 GB)</h3><p>One click downloads everything the six presets use:</p><ul><li>moss_image_sft_int8_convrot.safetensors (5.04 GB): the quality fine-tuned SFT model, the preset default.</li><li>moss_image_base_int8_convrot.safetensors (5.04 GB): the raw Base model for the Base presets, LoRA training and variety.</li><li>Qwen3.5 9B INT8 ConvRot text encoder (9.08 GB), saved as qwen3.5_9b_int8_convrot.safetensors, the file and name SwarmUI itself downloads for MOSS-Image, so SwarmUI reuses it.</li><li>FLUX 2 VAE (0.31 GB), saved as VAE/Flux/flux2-vae.safetensors, the same file FLUX 2 Klein and FLUX 3 Klein use.</li><li>MOSS_Image_SFT_DeBand_L37–41.safetensors (1.05 GB, MATLOWAI, Apache-2.0): the SFT banding fix that the presets apply automatically.</li><li>MOSS_Image_SFT_8_Steps_DMD_LoRA.safetensors (1.11 GB, lodestones, Apache-2.0): the 8-step LoRA of the Turbo preset, saved from lodestones’ latest release file, which can change upstream.</li></ul><p>Where to find it: run Windows_Start_Download_Models_App.bat. The MOSS-Image Core Bundle is the first bundle of the SwarmUI Bundles tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gqPJpJrTOzy4pnYc.png" /></figure><p>Optional, under Image Generation Models → MOSS-Image Models: the original BF16 SFT and Base models (9.47 GB each; BF16 Base is the recommended precision for LoRA training and fine-tuning), a smaller Qwen3.5 9B W4A8 GPTQ text encoder for low RAM and VRAM (6.84 GB; select it in the Qwen Model parameter), the FLUX 2 small-decoder VAE that decodes faster with less VRAM (0.23 GB; select it as the VAE) and an extra banding fix for stubborn smooth skies (1.05 GB).</p><p>Why INT8 ConvRot with PyTorch attention: on an RTX 3090 the INT8 ConvRot files run about twice as fast as BF16, and with PyTorch attention their images were close to the official BF16 outputs. The INT8 files’ own INT8 attention is about 20% faster, but it changed the images noticeably in our tests, so the presets use PyTorch attention. Our ComfyUI graph with these settings reproduced official MOSS-Image Space examples from the same seed, with the same composition, subject and details.</p><h3>UPDATED — AVATARFOREVER: LIVE PREVIEWS, TESTED LIP MOTION SETTINGS AND A HISTORY FIX</h3><ul><li>Live previews: the AvatarForever sampler now shows a preview frame after every denoising step, starting at step 1, and the finished video after export; in SwarmUI the previews appear in the normal generation preview. They use taeltx2_3.safetensors (0.02 GB) when you have it, now part of the LTX 2.3 Avatar Forever Core Bundle (44.53 GB), and a fast latent preview otherwise. Previews do not change the video: with previews on and off, the same 452-frame MP4 came out byte for byte in our test.</li><li>Tested lip motion settings: the AvatarForever Unified Image + Audio or Text + Audio preset now renders an exact 512×768 portrait canvas with 16 latent frames per chunk (was 4) and a prompt that asks for expressive mouth movement. In a six-second comparison with the same portrait, song, prompt and seed, 4-frame chunks kept the lips closed and 16-frame chunks gave visible mouth movement, which held through the complete 18-second song. At 1024×1024 the mouth stayed mostly closed in the same test even with 16-frame chunks, so the preset keeps 512×768. Larger chunks use more memory; check a short sample when you change the image, the audio or the resolution.</li><li>History fix (AvatarForever 1.0.3): an interrupted video export could leave an incomplete MP4 that kept SwarmUI’s History waiting, and with several browser tabs open even finished images stopped loading. SwarmUI now skips the metadata of such files; complete files and generation settings are unchanged.</li></ul><p>Where to find it: Presets tab → type AvatarForever → click the preset card → Apply, then attach a speech or song to the prompt and, if you like, a portrait as Init Image.</p><h3>NEW — MOSS-IMAGE PROMPT GUIDE FOR LLMS</h3><p>New file: Automatic_Prompt_Generate_Enhance_To_Video_And_Image_Models_LLM_Instructions/MOSS_Image_Enchance_Prompt_Feed_For_LLMs.txt. Give it to ChatGPT, Claude, Gemini or a local LLM together with your idea, and it writes MOSS-Image prompts: detailed captions, exact text for signs and posters, and layout drafts with boxes that the presets convert to MOSS-Image’s own coordinates at your output size. Its first part explains how the SwarmUI and ComfyUI presets use the prompt (recipe, prompt syntax, layout mode and resolutions); it is based on the official MOSS-Image user guide and its prompt enhancer.</p><h3>UPDATED — BUNDLE REPORT AND DOWNLOADER APP</h3><ul><li>Which_Bundles_Downloads_Which_Preset_Models.html is regenerated for V81: all 74 presets, and every MOSS-Image preset is fully covered by the MOSS-Image Core Bundle. The report generator now knows SwarmUI’s default MOSS-Image text encoder and VAE, the de-band LoRA that the MossImage extension loads by itself for SFT models, and AvatarForever’s live-preview file.</li><li>Windows_Start_Download_Models_App.bat now installs the downloader’s packages in pip’s isolated mode, so your user pip configuration and PIP_ option variables no longer affect its install.</li><li>The downloader app shows v183.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v183.zip from this post and extract it over your existing downloader folder.</li><li>Run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI to the version with native MOSS-Image support (added on 17 October), installs the new MossImage premium extension and AvatarForever 1.0.3, and updates SECoursesAudioTools.</li><li>Update the ComfyUI backend: MOSS-Image needs ComfyUI from 16 October or newer. If SwarmUI uses our ComfyUI installation, update it with our ComfyUI installer V143 (Windows_Install_Or_Update_ComfyUI.bat). If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start. You can switch it back afterwards. Our ComfyUI installer V143 also brings the same six MOSS-Image presets and the MOSS-Image prompt guide to ComfyUI.</li><li>Restart SwarmUI and its ComfyUI backend so the new extension and nodes register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v81.json. It backs up your current presets, thumbnails included, to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the MOSS-Image Core Bundle. AvatarForever users: also download taeltx2_3.safetensors (0.02 GB) for the live previews; it is in the LTX 2.3 Avatar Forever Core Bundle and under Video Generation Models → LTX 2.3 Video Models.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>16 October 2026 Update V182</h3><p>This is a FLUX 3 Klein update. V182 brings Black Forest Labs’ new FLUX 3 Klein image model to SwarmUI with three advanced presets and our new FLUX 3 Klein premium extension: text to image, image editing with up to 10 reference images, visual layout boxes, ComfyUI-style prompt syntax and prompt batches from .txt files. The model downloader gets the FLUX 3 Klein Core Bundle with our HQ INT8 ConvRot model (10.85 GB in total), and the zip adds a FLUX 3 Klein prompt-writing guide for LLMs.</p><p>Amazing SwarmUI Presets moves to V80 with 68 presets: 3 new FLUX 3 Klein presets, and the other 65 presets are unchanged. The new presets need the new FLUX 3 Klein Core Bundle (10.85 GB) of the model downloader.</p><p>All 68 presets as they appear in the SwarmUI Presets tab; the three FLUX 3 Klein presets are in the last row:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NGKg2SMOY8kLxO7S.jpeg" /></figure><p>Like every other preset, each FLUX 3 Klein preset has its own artwork that shows its task: Text To Image, Image Edit or Layout Boxes.</p><h3>NEW PRESETS — FLUX 3 KLEIN: TEXT TO IMAGE, IMAGE EDIT AND LAYOUT BOXES</h3><p><em>“FLUX 3 Klein — Advanced Text To Image — 261015”</em></p><p><em>“FLUX 3 Klein — Advanced Image Edit — 261015”</em></p><p><em>“FLUX 3 Klein — Advanced Layout Boxes — 261015”</em></p><p>All three load our HQ INT8 ConvRot FLUX 3 Klein model with SwarmUI’s own Qwen3-VL 4B text encoder and the FLUX 2 VAE, and set 1024×1024, 16 steps, Euler/simple and CFG 1 (CFG 1 ignores the negative prompt). Each preset turns on our FLUX 3 Klein extension in its own mode: Text / optional references, Reference edit or Layout / optional references. Measured in SwarmUI on an RTX 5090 at 1024×1024: about 2 seconds of generation time for an image and 3.9 seconds for an edit with one reference image.</p><p>Where to find it: Presets tab → type FLUX 3 Klein in the filter → click a preset card → Apply. Its settings fill the new FLUX.3 Klein group of the parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Eg6_qbjHDpMz5wYR.png" /></figure><p>The FLUX.3 Klein group and the preset settings in detail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CGd7-cNmrxEirnZR.png" /></figure><ul><li>Seed: -1 gives a new image every time; a fixed seed repeats the same image (pixel-identical in our tests).</li><li>Mode: text to image, reference edit or layout boxes. Reference images are optional in the text and layout modes.</li><li>Reference Megapixels: the size of each attached image, from 0.1 to 4 megapixels (1 is about 1024×1024), independent of the output size.</li><li>Your usual SwarmUI controls keep working: resolution, batch size, variation seed and compatible FLUX 3 Klein LoRAs.</li><li>The extension acts only on FLUX 3 Klein models. Apply any other preset and its own workflow runs unchanged.</li></ul><h3>NEW — EDIT IMAGES WITH UP TO 10 REFERENCE IMAGES (@image1, @image2)</h3><p>Attach images to the prompt with Add Image / Video / Audio and cite them in your prompt: @image1 is the first attachment, @image2 the second, and so on (@img1, @picture1 and @pic1 work too). Describe the complete result: what changes and what stays. The extension turns these names into FLUX 3 Klein’s own reference labels.</p><ul><li>Up to 10 images. Each keeps its aspect ratio, and transparent areas are filled with white.</li><li>Reference edit needs at least one image. If the prompt cites an image that is not attached, SwarmUI shows a clear error; video and audio attachments are refused with a message.</li></ul><p>Where to find it: apply the FLUX 3 Klein — Advanced Image Edit preset, attach your photo above the prompt and describe the result.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YmnsgzP1JztbKKh5.png" /></figure><p>Here the white mug of the attached photo became cobalt blue in 3.9 seconds of generation time, while the croissant, the window, the light and the camera angle stayed. To take one object out of several references, name the image it comes from and say which other objects should be absent: in our test with this coffee photo as @image1 and a red teapot as @image2, asking for the vessel from @image2 gave the teapot alone.</p><h3>NEW — VISUAL LAYOUT BOXES: PUT EVERY OBJECT WHERE YOU WANT IT</h3><p>Draw boxes on a canvas that has the shape of your output image, describe the object of each box, and cite the boxes in your prompt as &lt;obj_1&gt;, &lt;obj_2&gt;. SwarmUI turns them into FLUX 3 Klein’s native layout table, so the model places each object in its box.</p><ul><li>Draw a box with the mouse and drag it to move it; select a box and drag its lower-right corner to resize it.</li><li>Each box is an Object or Text / lettering with the exact text to render, and its edges can be typed in percent.</li><li>Add, delete and reorder up to 32 boxes, or open the JSON. Reference images still work in this mode.</li><li>Boxes guide the composition; they are not inpainting masks.</li></ul><p>Where to find it: apply the FLUX 3 Klein — Advanced Layout Boxes preset. The canvas is in the FLUX.3 Klein group.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*demf0HW-7W3Q87rn.png" /></figure><p>The preset’s two boxes put the red teapot on the left and the blue cup on the right, as drawn, in 2.0 seconds of generation time.</p><h3>NEW — COMFYUI-STYLE PROMPT SYNTAX</h3><p>The FLUX 3 Klein presets understand the prompt syntax of our ComfyUI presets, next to SwarmUI’s own prompt tags:</p><ul><li>(ceramic:1.1) weights a phrase.</li><li>[red:blue:0.5] switches from one word to the other halfway through sampling.</li><li>[red|blue] alternates the words from step to step.</li><li>&lt;break&gt; splits the prompt into separately encoded parts.</li><li>{red|blue} picks one at random, following the seed.</li><li>SwarmUI tags such as &lt;random:red|blue&gt;, &lt;fromto[0.5]:red|blue&gt; and &lt;alternate:red|blue&gt; keep working.</li><li>A complete native caption with its JSON layout table, as the new prompt guide writes it, works in the text and reference edit modes.</li></ul><h3>NEW — PROMPT BATCHES FROM .TXT FILES OR A WHOLE FOLDER</h3><p>Generate a series with the same settings and reference images. Add prompt adds a prompt you type; Import .txt and Import folder load text files, one complete prompt per file, in natural file-name order. Generate prompt batch runs your main prompt first and then every extra prompt; Stop after current ends the batch after the current prompt finishes. The normal Generate button still generates only the main prompt.</p><p>Where to find it: FLUX.3 Klein group → FLUX Klein Prompt List, in every FLUX 3 Klein preset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*H80n7Kj0tKvOaAs_.png" /></figure><p>Three imported .txt files plus the main prompt gave four images in 8 seconds, from click to finish. A batch takes up to 63 extra prompts.</p><h3>NEW — MODEL DOWNLOADER: FLUX 3 KLEIN CORE BUNDLE (10.85 GB)</h3><p>One click downloads everything the three presets use:</p><ul><li>FLUX-3-Image-Klein-INT8-ConvRot-HQ.safetensors (5.66 GB): our HQ INT8 ConvRot version of the distilled FLUX 3 Klein, with selected layers kept in BF16.</li><li>Qwen3-VL 4B FP8 text encoder (4.88 GB), saved as qwen3vl_4b.safetensors, the file name SwarmUI itself uses for FLUX 3 Klein.</li><li>FLUX 2 VAE (0.31 GB), saved as VAE/Flux/flux2-vae.safetensors, the same file FLUX 2 Klein uses: if you already have it, it is verified and skipped.</li></ul><p>Where to find it: run Windows_Start_Download_Models_App.bat. The FLUX 3 Klein Core Bundle is the first bundle of the SwarmUI Bundles tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1hAOX6QNNZPs6uo8.png" /></figure><p>Optional, under Image Generation Models → FLUX 3 Klein Models: the original distilled BF16 model (10.34 GB), the undistilled Base BF16 checkpoint for fine-tuning and LoRA training (10.32 GB; accept its license on Black Forest Labs’ gated Hugging Face page and enter your Hugging Face token in the downloader; it uses its own 50-step, guidance 3.5 settings), and a Qwen3-VL 4B BF16 text encoder (8.27 GB). The BF16 encoder has the same qwen3vl_4b.safetensors name as the FP8 one, so download only one of them.</p><p>Why HQ INT8: in our 40-state test against the original BF16 model (BF16 text encoder, RTX 3090), the denoiser output of our HQ INT8 ConvRot model differed by 2.14% on average, against 3.05% for the official INT8 ConvRot file with its default settings and 2.27% with matched attention; with matched attention, one held-out edit case favored the official file (3.84% against 3.58%; 4.73% with its default settings). The transformer peaked at about 6 GiB of VRAM instead of 10.5 GiB and ran each step about twice as fast as BF16. These are numerical-fidelity measurements, not image-quality percentages.</p><h3>UPDATED — TEXT ENCODERS GO WHERE YOUR SWARMUI LOOKS FOR THEM</h3><p>SwarmUI keeps text encoders in the folders of its SDClipFolder setting: text_encoders;clip on installs from late November 2025 on, and only clip on older installs, because SwarmUI never changes a setting it has saved. The downloader now reads this setting from your SwarmUI (Data/Settings.fds next to the Models folder you download into):</p><ul><li>A text encoder you already have in one of those folders is verified where it is, instead of being downloaded a second time.</li><li>A new text encoder goes into the first folder, where SwarmUI saves its own downloads too.</li><li>If the downloader cannot find your SwarmUI settings, it uses clip, which SwarmUI scans by default on every install.</li></ul><p>The ComfyUI and Forge folders are unchanged. We tested this on new, old and custom SwarmUI folder settings.</p><h3>NEW — FLUX 3 KLEIN PROMPT GUIDE FOR LLMS</h3><p>New file: Automatic_Prompt_Generate_Enhance_To_Video_And_Image_Models_LLM_Instructions/FLUX_3_Klein_Enchance_Prompt_Feed_For_LLMs.txt. Give it to ChatGPT, Claude, Gemini or a local LLM together with your idea, and it writes FLUX 3 Klein prompts: text-to-image captions, edits that say what stays and what changes, multi-reference prompts with @image1 and @image2, exact text for posters and signs, and layout boxes. Its first part explains the SwarmUI presets (modes, @image numbering, the box editor, prompt syntax and batches); the rest is the same research-based guide, version 1.5, as in our ComfyUI installer, with original examples.</p><h3>UPDATED — BUNDLE REPORT, PRESET BACKUPS AND DOWNLOADER APP</h3><ul><li>Which_Bundles_Downloads_Which_Preset_Models.html is regenerated for V80: all 68 presets, and the three FLUX 3 Klein presets are fully covered by the FLUX 3 Klein Core Bundle. The report generator now also knows SwarmUI’s default FLUX 3 Klein text encoder and VAE.</li><li>Preset backups keep their thumbnails: Windows_Preset_Delete_Import.bat now saves every preset’s thumbnail inside the backup file, so a backup holds everything needed to bring your own presets back with their images.</li><li>The downloader app shows v182 and uses Gradio 6.30.0. Windows_Start_Download_Models_App.bat updates Gradio automatically, and the RunPod / SimplePod and Massed Compute instructions install it.</li></ul><h3>TESTED — REAL GENERATIONS IN CHROME</h3><p>The extension and the presets were tested on 15 October with SwarmUI 967f005c and ComfyUI 20b46065 on an RTX 5090 in a real Chrome: text to image, scene-preserving edits, two-reference selection, visual layout (drawing, moving and resizing boxes), advanced prompt syntax, prompt batches, .txt file and folder import, Stop after current, and 23 workflow checks (reference numbering, the 10-image limit, duplicate images, layout tables, reference sizing and model gating). Every generated result shown in this post is a real run on the HQ INT8 model. The optional BF16 models and FLUX 3 Klein LoRAs were not generated separately.</p><h3>CUSTOM NODE PACKS WE DEVELOP OR MAINTAIN UPDATED</h3><p>SECoursesAudioTools</p><ul><li>LoRA and DoRA files trained with SECourses Musubi Trainer now load completely. ComfyUI silently skipped some of their layers: the MLP halves of Qwen Image 2.1, the q, k and v parts of Z-Image’s fused attention, and every HiDream-O1 layer. The node pack now maps the trainer’s layer names to them, DoRA scales included. SwarmUI gets this too: Windows_Install_or_Update_SwarmUI.bat installs this node pack for SwarmUI’s self-starting ComfyUI backend.</li><li>New FLUX.3 Gallery Prompt and FLUX.3 Optional Gallery References nodes connect our reference gallery to FLUX 3 Klein in our ComfyUI presets: @image1 names, seeded {red|blue} choices and up to 10 references with their own size.</li></ul><h3>SWARMUI PREMIUM EXTENSIONS UPDATED</h3><ul><li>FluxThreeKlein 1.0.0 (new): FLUX 3 Klein advanced prompting, image references, visual layout boxes and prompt batches on SwarmUI’s native FLUX 3 workflow. Windows_Install_or_Update_SwarmUI.bat installs it.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v182.zip from this post and extract it over your existing downloader folder.</li><li>Run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI to the version with native FLUX 3 Klein support (added on 15 October), installs the new FluxThreeKlein premium extension and updates SECoursesAudioTools.</li><li>Update the ComfyUI backend: FLUX 3 Klein needs ComfyUI from 15 October or newer (tested with 20b46065). If SwarmUI uses our ComfyUI installation, update it with our ComfyUI installer V141 (Windows_Install_Or_Update_ComfyUI.bat). If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start. You can switch it back afterwards.</li><li>Restart SwarmUI and its ComfyUI backend so the new extension and nodes register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v80.json. It backs up your current presets, thumbnails included, to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the FLUX 3 Klein Core Bundle.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><h3>10 October 2026 V180</h3><p>This is a MiniMax H3 RefMods, Long Lip Sync and presets update. V180 adds MiniMax H3 RefMods (save a character, a scene or a voice once and reuse it in every H3 preset), the new RefMod Builder preset, two MiniMax H3 Long Lip Sync presets (one complete soundtrack → one continuous video), new artwork for every preset, MiniMax H3 Optimizations on ComfyUI’s native SOL attention, model downloader bundles that now hold every model their presets use, and the matching updates of our SwarmUI premium extensions and custom nodes.</p><p>Amazing SwarmUI Presets moves to V79 with 65 presets: 3 new presets, the 4 Wan 2.1 presets removed, and all 62 other presets updated. No new model downloads are needed for the SwarmUI presets. The new presets use the models of the existing MiniMax-H3 Core Bundle.</p><p>All 65 presets as they appear in the SwarmUI Presets tab, with their new thumbnails:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GltDO3betHPaaRNR.jpeg" /></figure><p>Every preset now has its own artwork: a different animal or mythical creature for each preset, with the model family or task and the mode or step count on the thumbnail, so you can find the right preset at a glance.</p><p>Exact V178 → V179 zip audit:</p><ul><li>7 changed files: Amazing_SwarmUI_Presets_v77.json was replaced by Amazing_SwarmUI_Presets_v79.json, the downloader app shows the new version, update_swarmui.py now moves SwarmUI’s ComfyUI-GGUF and ComfyUI-LTXVideo node packs to our fixed forks, and the model catalog, its size data, the bundle report generator and the regenerated Which_Bundles_Downloads_Which_Preset_Models.html carry the bundle updates described below.</li><li>The other 35 files are identical.</li><li>Everything else for SwarmUI arrives through Windows_Install_or_Update_SwarmUI.bat: our premium extensions, FoleyExtension, SECoursesAudioTools, ComfyUI-TeaCache and the ComfyUI-GGUF and ComfyUI-LTXVideo fork switch. Our ComfyUI installer V139 brings the same fork fixes plus ComfyUI-WanVideoWrapper.</li></ul><h3>NEW — MINIMAX H3 REFMODS: SAVE A CHARACTER, SCENE OR VOICE ONCE, REUSE IT IN EVERY H3 PRESET</h3><p>A RefMod is a small .safetensors file that stores ready-encoded MiniMax H3 references: photos, a video with its soundtrack, audio, or a mix. Build it once, then cite it in any MiniMax H3 prompt as @refmod1, @refmod2 and so on. You no longer attach the same media and VAE-encode it again for every generation.</p><ul><li>Works in 11 MiniMax H3 presets: text to image, text to video, image to video, first and last frame, references to video, audio only, and both new Long Lip Sync presets.</li><li>No attachment needed: RefMods alone are enough for MiniMax H3 References, and normal prompt attachments still work next to them.</li><li>Every row has its own controls: enable, strength from 0 to 1 (how much reference detail is kept) and a component filter: all, visual or audio. Use only the voice of a video bundle, or only its look.</li><li>The right model is picked for you: active RefMods switch an FL2VA model to the matching Ref2VA model, and the FL2V Turbo 4-step and 8-step LoRAs to their Ref2V versions.</li><li>No fixed number of files: add as many rows as your VRAM and context allow. The picker shows the extra reference tokens.</li><li>Fully optional: empty, disabled and zero-strength rows change nothing, and RefMods are off in every preset until you turn them on.</li><li>Compatible: reads MiniMaxH3Mod v4 and legacy single-reference files, v5 bundles and Fantastic files, and new bundles open in those projects too. Files live in the ComfyUI backend’s models/refmods folder (SwarmUI’s Models/refmods is found as well); Import files uploads them from your computer and never overwrites an existing file.</li></ul><p>Where to find it: apply any MiniMax H3 preset, open the MiniMax H3 RefMods group (between Resolution and Sampling) and tick MiniMax H3 RefMods.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5NE1GIK_HcmSFDtW.png" /></figure><p>Each row picks a file, its strength and which parts of the bundle to use:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g7NRXKHjNSed5jYT.png" /></figure><p>Here @refmod1 is a character-and-voice bundle at full strength, and @refmod2 adds only the voice of a video bundle at strength 0.8. Validated on an RTX 5090: all 11 RefMods-enabled MiniMax H3 presets generated in SwarmUI with active RefMods, and in our parity check, disabled and zero-strength rows built exactly the same SwarmUI workflow as no RefMods. RefMods add reference conditioning; they do not guarantee an exact identity or voice clone.</p><h3>NEW PRESET — MINIMAX H3 REFMOD BUILDER: BUILD YOUR OWN REFMODS INSIDE SWARMUI</h3><p><em>“MiniMax H3 RefMod Builder — Images Video Audio — 261009”</em></p><p>Apply the preset first (it clears the prompt, Init Image and Init Audio), then attach photos of one character, a video, audio or a mix above the prompt, set the RefMod name and press Generate. Only the H3 video and audio VAEs run: no diffusion sampling, no text encoder and no training. On an RTX 5090 a bundle took a few seconds in our tests (4.6 seconds of generation time in SwarmUI). The file is saved in models/refmods and a source preview returns to SwarmUI; press Refresh in the RefMod picker to use it right away.</p><ul><li>One character stacks several photos into one visual reference; separate references keeps them as numbered pictures.</li><li>Videos keep their soundtracks, so one bundle can carry both a look and a voice.</li><li>Resolution 512 and compression none are the defaults. Compression 32, 16 or 8 pools the encoded references spatially to save tokens.</li><li>Existing files are never overwritten: new files get numbered names. One bundle can hold up to 256 references.</li></ul><p>Where to find it: Presets tab → MiniMax H3 RefMod Builder. Its settings are in the MiniMax H3 RefMods group.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ihXuRg3WYlogdZuk.png" /></figure><p>The builder settings in detail:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OX2cKNmR_oGvrZqK.png" /></figure><p>Here three photos of the same presenter become one “my_presenter” character RefMod.</p><h3>NEW PRESETS — MINIMAX H3 LONG LIP SYNC: ONE COMPLETE SOUNDTRACK → ONE CONTINUOUS VIDEO</h3><p><em>“MiniMax H3 Long Lip Sync — References — Turbo 4–261008”</em></p><p><em>“MiniMax H3 Long Lip Sync — First Frame — Turbo 4–261008”</em></p><p>Put your complete narration or song in Init Audio, add a portrait and press Generate. The video follows the whole recording automatically: there is no manual duration, and every word and pause keeps its original timing. The result is one 24 FPS H.264/AAC MP4 with your soundtrack.</p><ul><li>Native latent continuation: the video is generated in overlapping native H3 windows that follow your soundtrack (default 243 frames, about 10.1 seconds) with a 39-frame overlap (about 1.6 seconds). Each window continues from the previous window’s clean latents instead of re-encoding decoded pixels.</li><li>Two modes: References attaches the portrait beside the prompt as @image1 (Ref2VA); First Frame puts it in Init Image (FL2VA).</li><li>Tuned defaults: 832×1248, Turbo 4 steps and MiniMax H3 Low VRAM on, with MiniMax H3 Optimizations and First Block Cache off. Long mode always samples with Euler/simple at CFG 1, so the Sampler, Scheduler and CFG Scale settings are not used.</li><li>Video Face Inpainting must be off in long mode; SwarmUI shows a message if it is on.</li><li>Adjustable: H3 Long Window Frames 141, 192, 243, 294 or 345 and H3 Long Overlap Frames 39 or 90. Smaller windows need less memory.</li><li>The prompt token meter shows the estimate for one window, so a long soundtrack stays within budget.</li><li>RefMods work here too.</li></ul><p>Where to find it: Presets tab → a Long Lip Sync preset. The new controls are in the Init Audio group.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*234Ybq_I3XQUUF1Q.png" /></figure><p>The new Init Audio controls, here with a 2-minute narration:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aB3MR8LLFC8ss9k4.png" /></figure><p>Measured in SwarmUI end to end on an RTX 5090: a 30-second 512×768 First Frame video took 125.81 seconds, and a 60-second 832×1248 References video took 753.98 seconds (with MiniMax H3 Optimizations on), both covering the full audio. Long mode is experimental, so review every result: joins can show short soft frames or motion changes, and very long high-resolution clips can drift in framing.</p><h3>UPDATED — THE 9 EXISTING MINIMAX H3 PRESETS ARE REFMODS-READY</h3><p>The 9 existing MiniMax H3 presets now include the optional RefMods settings, explicitly switch Long Lip Sync, RefMods and the RefMod Builder off, and set MiniMax H3 Video Shift to 12, the value they already used. Applying one of them after a Long Lip Sync or RefMod Builder preset therefore turns those features off and restores Video Shift 12 (the First Frame preset uses 6). Their descriptions explain how to use @refmod1, @refmod2.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R-dMDy0l6flkNZLd.png" /></figure><ul><li>Updated: Text to Image with Optional References, References to Video, Text to Video, Beginning and Ending Image/Frame to Video, Image to Video, Audio Only with Optional References, Text to Video 8 Steps, Beginning and Ending Image/Frame to Video 8 Steps, and References to Video 8 Steps.</li><li>New: Long Lip Sync — References, Long Lip Sync — First Frame and the RefMod Builder.</li></ul><h3>REMOVED — THE 4 WAN 2.1 PRESETS</h3><ul><li>Removed from Amazing SwarmUI Presets: Wan 21 Text To Video Base Model, Wan 21 FusionX Text to Video, Wan 21 Image To Video Base Model and Wan 21 LightX2V LoRA Image to Video. Windows_Preset_Delete_Import.bat removes them when it imports V79.</li><li>The Wan 2.1 models and the Wan 2.1 Core Models Bundle stay in the model downloader, and the Wan 2.2 presets are unchanged.</li></ul><h3>UPDATED — NEW ARTWORK FOR EVERY PRESET AND TWO PRESET FIXES</h3><ul><li>New thumbnails for all 65 presets, shown in the first image of this post.</li><li>FLUX 2–8 Steps Fast Realism — Low RAM — 251230: Flux Guidance Scale 7 → 1. It now matches the standard Fast Realism preset; only the GGUF model and the GGUF text encoder differ.</li><li>FLUX Kontext Dev Edit Images: the description now explains how to use it: put the image in Init Image (or paste it into the prompt) and describe the change. SwarmUI detects the official Kontext file as plain Flux.1 Dev, which ignores your image, so the description also gives the one-time fix: the model’s ☰ menu → Edit Metadata → Architecture: Flux.1 Kontext Dev → Save.</li></ul><h3>UPDATED — MINIMAX H3 OPTIMIZATIONS NOW USE COMFYUI’S NATIVE SOL ATTENTION</h3><p>The MiniMax H3 speed option is now called MiniMax H3 Optimizations (it was MiniMax H3 4x Speed). Its parameter ID did not change, so saved presets and API calls keep working. Through our ComfyUI-TeaCache fork it now uses ComfyUI’s own native H3 sparse attention (SOL), and SOL and caching have separate switches:</p><ul><li>MiniMax H3 Speed Sparse Attention (SOL): auto benchmarks the complete attention path on your GPU and keeps SOL only when it is at least 5% faster; enabled skips the benchmark; disabled keeps normal attention.</li><li>MiniMax H3 First Block Cache (new switch): on by default, as before. Untick it to run SOL alone.</li><li>First Block Cache now works together with other native attention patches, such as SLA, instead of replacing them.</li></ul><p>Where to find it: type MiniMax H3 into Filter parameters, tick Display Advanced Options? and tick MiniMax H3 Optimizations in Core Parameters. First Block Cache then appears in Advanced Sampling next to Speed Sparse Attention; both apply only while Optimizations is on.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eE3Ns7id5xzU3Y-A.png" /></figure><p>The same controls in the parameter panel, numbered like the steps above:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Boko0RR-ejeTb6m8.png" /></figure><p>Measured on an RTX 5090 (FL2VA INT8, Turbo 4 steps, 832×1248, 10-second narration, First Block Cache off): 110.9 s with dense attention and 99.2 s with the default protected SOL schedule (1.12×); at 512×768 the gain was 1.06×. In a matched run, the old bundled integration took 207.1 s against 116.0 s dense, so this update removes that slowdown. SOL and caching are approximations, not lossless.</p><h3>FIXED — INIT AUDIO MATCH DURATION IN SWARMUI’S OWN MINIMAX H3 GRAPHS</h3><p>With Init Audio and Init Audio Match Duration on (the default), the video length now follows your audio in text to video, image to video with a MiniMax H3 Video Model, and native reference generations. These graphs used to keep the Text2Video Frames or Video Frames length, so longer audio was cut and shorter audio was padded with silence. Now, for example, 2.5 to 3 seconds of audio gives 73 frames (3.04 seconds), tested at 22.05, 44.1 and 48 kHz.</p><h3>FIXED — OPTIONS OF DISABLED FEATURES ARE NO LONGER SENT WITH EVERY GENERATION</h3><p>The option groups of our extensions (AvatarForever, Qwen Unified, SVI Pro, LTX 2.3 Foley, LTX 2.5 Audio To Video, LTX 2.3 Licon MSR and the MiniMax H3 groups, including Video Face Inpainting) stayed visible and were sent with every generation, even while their feature was off. In one test, a plain Z-Image Turbo picture carried 85 unused parameters, and SwarmUI SHA-256 hashed 11 unrelated model files (47.1 GB) into its metadata. Now these options stay hidden while their feature is off and are no longer sent: the same picture has 0 unused parameters and is pixel-identical. With a feature on, its values are sent exactly as before.</p><h3>FIXED — MINIMAX H3 AUDIO NODES WORK WITHOUT TORCHAUDIO</h3><p>ComfyUI removed torchaudio on 21 September, and new Linux SwarmUI installs no longer have it. Our MiniMax H3 Init Audio node (FoleyExtension) and the H3 Long Lip Sync and H3 Streaming nodes (SECoursesAudioTools) now resample with ComfyUI’s own resampler, so any sample rate works on a torchaudio-free install. The new resampling was bit-identical in all 88 comparisons, and a real MiniMax H3 Turbo 4 run with a 48 kHz soundtrack gave 120 of 120 identical frames and bit-identical audio.</p><h3>UPDATED — MODEL DOWNLOADER: EVERY CORE BUNDLE NOW DOWNLOADS ALL MODELS ITS PRESETS USE</h3><p>We checked all 65 SwarmUI presets and all 90 presets of our ComfyUI installer against the downloader: every model file a preset names, including bypassed nodes and optional switches, and every model our extensions and nodes load by default, is now downloaded by that preset’s own core bundle. Files that a node downloads by itself, such as the RIFE interpolation model, are not part of the bundles.</p><ul><li>MiniMax-H3 Core Bundle (67.55 → 69.91 GB) and MiniMax-H3 Low VRAM Bundle (51.85 → 54.21 GB): add codeformer.pth and the InsightFace det_10g.onnx detector for the CodeFormer mouth pass of the ComfyUI MiniMax H3 Lip Synch preset, the InsightFace recognition model (w600k_r50.onnx) that MiniMax H3 face refinement uses to tell faces apart and the landmark model (2d106det.onnx) of its landmark-align node, and the LightX2V Turbo-SLA 4-step LoRA that the ComfyUI H3 lip sync presets load when SLA sparse attention is on.</li><li>LTX 2.3 Avatar Forever Core Bundle (44.34 → 44.50 GB): adds w600k_r50.onnx and 2d106det.onnx next to det_10g.onnx. insightface downloads its buffalo_l pack only when that folder does not exist yet, so a folder with det_10g.onnx alone left the face refinement without its recognition and landmark models. The updated face refinement also repairs such a folder itself.</li><li>LTX 2.5 Video Core Bundle (64.64 → 64.82 GB): adds the same three InsightFace files, which the LTX 2.5 Audio To Video presets use for face detection (without them they fall back to a simpler detector).</li><li>New SDXL Core Bundle (6.46 GB): RealVisXL V5, the model of the ComfyUI SDXL preset, which no bundle downloaded before.</li><li>Which_Bundles_Downloads_Which_Preset_Models.html is regenerated for V79. The report now reads the model settings of our AvatarForever and Qwen Image 2.1 extensions and SwarmUI’s default MiniMax H3 and LTX 2.5 text encoders and VAEs, so every preset shows its own core bundle with full coverage. Run Windows_Update_And_Open_Model_Report.bat to check your own model folders.</li><li>The MiniMax-H3 Low VRAM Bundle holds the W4A8 versions of the two main H3 models: with only that bundle, select them in place of the INT8 models the presets name.</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download SwarmUI_Model_Downloader_v179.zip from this post and extract it over your existing downloader folder.</li><li>Run Windows_Install_or_Update_SwarmUI.bat. It updates SwarmUI, our premium extensions (MiniMax H3 References 1.19.1, AvatarForever 1.0.2), FoleyExtension, SECoursesAudioTools and our ComfyUI-TeaCache fork, and it moves an existing ComfyUI-GGUF (SwarmUI’s DLNodes) and ComfyUI-LTXVideo (SwarmUI’s own ComfyUI) to our fixed forks. If you install GGUF support in SwarmUI later, run this .bat once more.</li><li>Update the ComfyUI backend: native SOL needs comfy-kitchen 0.2.37 or newer, and Long Lip Sync was tested with ComfyUI 52f98af2 (v0.39.0 plus 24 commits, 8 October). If SwarmUI uses our ComfyUI installation, update it with our ComfyUI installer V139: Windows_Install_Or_Update_ComfyUI.bat, plus Windows_Custom_Nodes_Bundles_Installer.bat options 5 and 17 for ComfyUI-LTXVideo and ComfyUI-WanVideoWrapper. It updates ComfyUI, comfy-kitchen and our forks together; updating ComfyUI alone breaks the GGUF and LTX 2.3 Foley presets. If SwarmUI uses its own ComfyUI, open Server → Backends, edit the ComfyUI Self-Starting backend and set AutoUpdate to Always Update for one start (step 2 already moved its GGUF and LTXVideo node packs to our forks). You can switch it back afterwards.</li><li>Restart SwarmUI and its ComfyUI backend so the new nodes and parameters register.</li><li>With SwarmUI running, run Windows_Preset_Delete_Import.bat to import Amazing_SwarmUI_Presets_v79.json. It backs up your current presets to utilities\presets_backups, deletes them all (your own presets too) and imports the new pack.</li><li>Run Windows_Start_Download_Models_App.bat and download the MiniMax-H3 Core Bundle (or the Low VRAM Bundle). If you already have it, run it again: files that are already there and verified are skipped, so only the new files download.</li></ol><p>ComfyUI Installer and Presets: <a href="https://www.patreon.com/posts/105023709">https://www.patreon.com/posts/105023709</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=847390a8fab0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Latest ComfyUI Updates — Lots of New Models, Improvements and Presets]]></title>
            <link>https://medium.com/@furkangozukara/latest-comfyui-updates-lots-of-new-models-improvements-and-presets-ca62298f4176?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/ca62298f4176</guid>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 10:28:01 GMT</pubDate>
            <atom:updated>2026-10-23T10:28:01.340Z</atom:updated>
            <content:encoded><![CDATA[<h3>Download installer zip file from here : <a href="https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709">https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709</a></h3><h3>Here below latest 4 updates we made</h3><h3>23 October 2026 Update V144</h3><p>This is a Fibo 2.0, MOSS Image Edit and FastH3 Omni update. V144 adds ten new presets: four for Bria’s Fibo 2.0, the new open image model that generates and edits images from detailed JSON captions (Dev with the official 50-step recipe and 4-step Turbo, each for text to image and image editing); two MOSS Image Edit presets that turn MOSS-Image into an instruction-based image editor (official 30 steps and Turbo 8 steps); and four presets for FastH3 Omni, MiniMax H3 distilled to 8 steps (References To Video, Text With References To Image, Audio Only and Long Lip Synch). It also adds a Fibo 2.0 prompt guide for ChatGPT, Claude and Gemini, updates the MOSS-Image and MiniMax H3 prompt guides and the two Qwen Image 2.1 presets, and brings the newest ComfyUI with native Fibo 2.0 and FastH3 Omni support plus the matching custom-node and SwarmUI extension updates.</p><p>Four native 1920×1088 images from the new Fibo 2.0 Dev Advanced Text To Image preset, with no upscaling: JSON captions, 50 steps, 36 to 38 seconds each on an RTX 5090. Every requested sign, label, menu line and poster line came out exactly as written.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r-fxgB3Q0XNd38xG.jpeg" /></figure><p>Every result in this post comes from a real run of the shipped presets in ComfyUI started with our Windows_Run_ComfyUI_Default.bat: each preset was opened from the Workflows sidebar and run with the Run button. Only the prompt, the input images or audio, the seed, the output size and the duration were changed; the MOSS Image Edit result in the edit comparison also used a larger reference size, as noted there. The times are ComfyUI’s generation times on our RTX 5090; a first run includes loading the models.</p><p>Exact V143 → V144 zip audit:</p><ul><li>11 new files: the 10 new presets and the Fibo 2.0 LLM prompt guide.</li><li>5 changed files: the MOSS-Image and MiniMax H3 prompt guides, the two Qwen Image 2.1 presets and Windows_Install_Or_Update_ComfyUI.bat.</li><li>No files were removed. The other 223 files are identical. Fibo 2.0 and FastH3 Omni support arrive through the installers: the ComfyUI update and our updated SECoursesAudioTools, ComfyUI-TeaCache and FoleyExtension.</li></ul><p>Here is the complete Presets folder of V144: 106 ready-to-use ComfyUI presets, plus 3 API presets in the api_paid folder. The ten new presets are marked NEW and the two updated Qwen Image 2.1 presets are marked UPDATED:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3yYU9XfGPDx08hKd.png" /></figure><p>The presets cover image, video, lip sync, audio and music models: AvatarForever, ERNIE Image, FastH3 Omni, Fibo 2.0, FLUX.1 Dev, Kontext and SRPO, FLUX 2, FLUX 3 Klein, Ideogram 4, Krea 2, LTX 2, LTX 2.3, LTX 2.5, MiniMax H3, MiniMax Music 3, MOSS-Image and MOSS Image Edit, Phantom Wan, Qwen Image, SDXL, Stable Video Infinity 2 Pro, Wan 2.1, Wan 2.2 and Z-Image.</p><p>Here are the new and updated presets, with what each one does:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CC3WhSl7uRO8dlGA.png" /></figure><p>The sections below explain each new model and preset with real results, then the updated Qwen Image 2.1 presets, prompt guides, node packs and the ComfyUI update.</p><h3>NEW PRESETS — Fibo 2.0: Text To Image and Image Edit, Dev 50 Steps and Turbo 4 Steps</h3><p><em>“Fibo 2.0 Dev — Advanced Text To Image — 261022”</em></p><p><em>“Fibo 2.0 Dev — Advanced Image Edit — 261022”</em></p><p><em>“Fibo 2.0 Turbo — Text To Image 4 Steps — 261022”</em></p><p><em>“Fibo 2.0 Turbo — Image Edit 4 Steps — 261022”</em></p><p>Fibo 2.0 is Bria’s new open 8B image model. One model generates and edits: text to image, instruction edits, masked edits and edits that combine up to 4 images. It was trained on long structured JSON captions (Bria’s VGL format) read by the Qwen3-VL 4B text encoder, and it uses the FLUX 2 VAE. A JSON caption controls every object, its position, the lighting, the camera and each piece of visible text; plain text works too. Our new Fibo 2.0 prompt guide makes ChatGPT, Claude or Gemini write the JSON for you.</p><p>Dev and Turbo: Dev is the full model with CFG and gives the best quality with the official recipe (50 steps, CFG 5). Edit Turbo is Bria’s distilled editing model (4 steps, CFG 1); Bria lists it for editing only, and in our tests it also makes very good 4-step text to image. Both are INT8 ConvRot files in our Fibo 2.0 Core Bundle.</p><p>All four presets share the same easy controls:</p><ul><li>One prompt box: a JSON caption or plain text. With no images in the gallery it makes a new image; add 1 to 4 images with + Add images to edit (or drop or paste them) and the prompt becomes the edit instruction, with @image1, @image2 for each image. + Prompt adds more prompts to one run, a folder of .txt files runs a whole prompt list, and Images per prompt makes several images of each prompt.</li><li>Output size for text to image: an aspect ratio and megapixels, or width and height. About 1 megapixel is the trained size and 2 megapixels works well.</li><li>Edits keep the photo’s own size: a 2 MP photo is edited at 2 MP, anything from 1 to 4 MP keeps its size, larger photos are scaled down to 4 MP and smaller ones go to the trained 1 MP size. Choose “Bria default” to edit everything at about 1 MP, as ComfyUI’s own node does.</li><li>Masked edit: load a photo in 06 OPTIONAL MASKED EDIT, click Paint mask (native editor), paint the area to change and describe what goes there. Every pixel outside the mask is copied back from your photo at the edit size.</li><li>Bria’s official sampling: the shift follows the image size like Bria’s own pipeline (ComfyUI uses the 1 MP value at every size), and Dev text to image uses Bria’s JSON negative prompt from the model card; edits leave the negative empty, which ComfyUI encodes as the null caption (a JSON caption with every field empty), like Bria’s edit negative.</li><li>Prompt syntax: random choices {red|blue} follow the seed in plain-text prompts; weights like (word:1.2) are removed because Qwen3-VL would read them as words; JSON is passed on unchanged (ComfyUI’s encoder only drops empty fields, as in training).</li><li>Seed fixed or randomized, and an optional LoRA stack, off by default. With Comfy-Org’s Fibo 2.0 turbo LoRA (an optional download) in the LoRA stack, Dev also works at 4 steps and CFG 1.</li></ul><p>Here is the Dev Advanced Text To Image preset right after the real run of the ramen-stall caption from the top image:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r_6A15CDHXPXLnxS.png" /></figure><p>The 1920×1088 image took 37.5 seconds. The first caption asked for glowing shop signs in the alley, and Fibo 2.0 filled them with invented lettering; replacing that phrase with “dark, unlit shop fronts and blank signboards” and “No other readable text anywhere” removed almost all of it, a good example of how exactly Fibo 2.0 follows a JSON caption.</p><p>Image edits: here is the same instruction through both Fibo 2.0 edit presets and the new MOSS Image Edit preset: “Change the scene to a snowy winter evening: fresh snow on the roof, the path and the grass, softly falling snow, and warm light glowing in the windows. Keep the building, the TIDE &amp; PAGE BOOKS sign and the chalkboard text exactly the same.” The original is the 1920×1088 bookshop from the top image, so the edits kept its 2 MP size:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vSS_8c21Ts4Ff7OL.jpeg" /></figure><ul><li>Fibo 2.0 Dev Advanced Image Edit (50 steps, CFG 5): 87.8 seconds at 2 MP. The closest to the original: the building, the sign, the chalkboard and even the sleeping cat stay, now under fresh snow.</li><li>Fibo 2.0 Turbo Image Edit 4 Steps (4 steps, CFG 1): 4.6 seconds at 2 MP, about 19 times faster, with nearly the same result. At 1 MP an edit took 3.4 seconds, including loading the model.</li><li>MOSS Image Edit Advanced (30 steps, CFG 4, reference size raised from 1024 to 1920 so that it also edits at 2 MP): 35.6 seconds. It kept the sign and chalkboard text but changed more of the scene around the shop.</li></ul><h3>NEW PRESETS — MOSS Image Edit: Official 30 Steps and Turbo 8 Steps</h3><p><em>“MOSS Image Edit — Advanced Image Edit — 261022”</em></p><p><em>“MOSS Image Edit — Turbo 8 Steps — 261022”</em></p><p>ChenkinLab’s Moss-Image-Edit is a LoRA for MOSS-Image SFT that edits images from a text instruction and up to 4 images. The LoRA alone does not edit: it also needs a reference branch that ChenkinLab ships in its own runtime. Our new MOSS-Image Edit References node adds that branch to ComfyUI’s native MOSS-Image model, so these presets run the official recipe with the MOSS-Image models you already have. The newest ComfyUI also has an experimental native version; these presets use our node.</p><ul><li>Advanced Image Edit: the official recipe: MOSS-Image SFT with the edit LoRA, 30 steps, Euler / normal, CFG 4, an empty negative and the automatic SFT de-band.</li><li>Turbo 8 Steps: adds the 8-step LoRA at 0.7 (8 steps, CFG 1): about 3 seconds per edit, close to the 30-step preset in our season, style and simple background edits. A completely new background behind a busy photo is harder: in a SwarmUI test with the same recipe, Turbo kept a busy street on all 5 seeds where the 30-step recipe made a beach, so start big background changes with Advanced.</li><li>How to use: add the image to edit to the gallery and write the instruction: what changes, then what stays. With several images, image 1 is the one edited and the others are references; name them “image 1”, “image 2” (or @image1, @image2).</li><li>Result size = image 1’s size, up to 1024 pixels on the long side; raise reference_max_edge to 1536 for bigger results. The result lines up with image 1 token by token, which is why it keeps the layout and identity.</li><li>Same prompt syntax, seed, LoRA stack and folder batch as the MOSS-Image presets of V143.</li></ul><p>Here is the Advanced Image Edit preset after a real run with its shipped settings. The image is the presenter portrait from the FastH3 Omni section below, with the instruction “Change her mustard yellow sweater to a dark green raincoat and add round tortoiseshell glasses. Keep her face, hair, smile and the plant-filled studio exactly the same.”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cdalTLpxyje9tg0M.png" /></figure><p>Her face, hair, smile and the plant-filled studio stayed the same; the 688×1024 result took 11.6 seconds.</p><h3>NEW PRESETS — FastH3 Omni: 8-Step Video, Image, Audio and Long Lip Sync</h3><p><em>“FastH3 Omni — References To Video 8 Steps — 261022”</em></p><p><em>“FastH3 Omni — Text With References To Image 8 Steps — 261022”</em></p><p><em>“FastH3 Omni — Audio Only With Optional References 8 Steps — 261022”</em></p><p><em>“FastH3 Omni — Long Lip Synch — Audio + Image 8 Steps — 261022”</em></p><p>FastH3 Omni, by FastVideo and Nuva Lab, is MiniMax H3 Ref2VA distilled to 8 steps: one model makes video with synchronized 32 kHz stereo sound from text plus image, video and audio references. It needs no turbo LoRA; the official recipe (8 steps, Euler / simple, CFG 1, video shift 12, audio shift 3) is already set. It was trained with Video Sparse Attention (VSA), and the video presets run that trained pattern through our updated speed node: sampling about 1.4 times (5-second clip) to 1.9 times (10-second clip) faster than dense attention at 1344×768, with the same quality in our tests.</p><ul><li>References To Video: 1344×768, 5 seconds by default (longer works). Add up to 12 references (at most 9 images, 3 videos and 3 audio files) and mention them as @image1, @video1, @audio1. Optional INIT IMAGE = the exact first frame, optional INIT AUDIO = a soundtrack the video must follow (lip sync, timing). Folder batches, merged videos and an optional face refine are included.</li><li>Text With References To Image: 1920×1080 stills from text plus up to 9 reference images; the model renders a short frame packet and the first frame is saved as the PNG.</li><li>Audio Only: speech, music or sound effects from text, with optional voice, music or video references; saves a FLAC file.</li><li>Long Lip Synch: your soundtrack plus a portrait gives one continuous talking video as long as the audio, made in overlapping 8-step windows of about 10 seconds that hold 39 frames at every join. An optional transcript of the exact words and an optional mouth restoration pass are included.</li><li>The speed node is on in the two video presets and off in Text With References To Image and Audio Only. A Low VRAM mode, an optional LoRA stack and our RefMods are in every preset.</li></ul><p>Here is what FastH3 Omni made from one reference. We made a presenter portrait with Fibo 2.0, used it as @image1 in Text With References To Image and References To Video, made 10 seconds of audio with Audio Only from a text prompt that describes her voice and her line (14.0 seconds), and gave that audio and the portrait to Long Lip Synch:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YI1lPmqBTqdQWdx_.jpeg" /></figure><ul><li>Text With References To Image: she became a pastry chef in a new scene, with her face, hair and freckles kept: 1920×1080 in 18.9 seconds.</li><li>References To Video: a 5-second 1344×768 clip with sound, prompted to show her walking through a greenhouse and speaking to the camera: 56.0 seconds.</li><li>Long Lip Synch: the 10-second audio from Audio Only plus the portrait gave a 10-second 832×1248 talking video in 112.0 seconds.</li></ul><p>Here are the controls of References To Video, with the portrait as @image1 and the structured prompt our MiniMax H3 prompt guide writes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vlpXimUHkvlnek1H.png" /></figure><ul><li>Click Add references on the big prompt node, or drag, drop or paste media onto it. Every card shows its token, and typing @ in the prompt inserts one.</li><li>The token meter under the prompt estimates how much of the model’s context the run uses and which mode it runs in, here “reference to video”.</li><li>Optional face refine (off by default) regenerates small faces and blends them back; see the fixes below.</li></ul><p>And here is Long Lip Synch after the real run with the generated audio and the portrait:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Mutx792MoJuDHT4l.png" /></figure><ul><li>Choose your soundtrack in 1. FULL SOUNDTRACK; the video is as long as the audio. Pick the portrait in 2. OPTIONAL IMAGE (none = audio and text only) and describe the scene in 3. SCENE.</li><li>The preset still points to our demo audio and image file names; choose your own files the first time you open it.</li><li>On an RTX 5090, 15 seconds of audio (two windows) at 832×1248 take about 3.5 minutes. Smaller windows (141 or 192 frames) use less VRAM.</li></ul><h3>UPDATED PRESETS — Qwen Image 2.1: Text Encoder File Name</h3><p><em>“Qwen Image 2.1 — Unified Generate Edit Inpaint RGBA — 260920”</em></p><p><em>“Qwen Image 2.1 — Unified Generate Edit Inpaint RGBA Ultimate Quality — Slow — 260922”</em></p><p>Both presets now load the Qwen Image 2.1 text encoder as qwen3vl_8b_int8_convrot.safetensors, its original name. Our downloader used to save it as qwen3vl_8b.safetensors, the name SwarmUI uses for Ideogram 4’s different encoder, so SwarmUI downloaded 9.35 GB again for Qwen Image 2.1. The updated downloader (SwarmUI_Model_Downloader_v184.zip) saves it under the original name, and both presets were tested with it.</p><p>If you downloaded Qwen Image 2.1 or Ideogram 4 with an older downloader: it saved all three Qwen Image 2.1 encoders and both Ideogram 4 encoders as text_encoders/qwen3vl_8b.safetensors, so the last download won. Check that file’s size (SwarmUI may have added about 64 KB of metadata to a file it used):</p><ul><li>9,350,798,360 bytes: the Qwen Image 2.1 INT8 ConvRot encoder; rename it to qwen3vl_8b_int8_convrot.safetensors.</li><li>17,534,334,616 bytes: the BF16 encoder; rename it to qwen3vl_8b_bf16.safetensors.</li><li>6,312,105,364 bytes: the W4A8 encoder; rename it to qwen3vl_8b_w4a8.safetensors.</li><li>10,588,637,512 or 6,305,221,764 bytes: Ideogram 4’s own FP8 or NVFP4 encoder; keep it as it is.</li></ul><p>After a rename, Ideogram 4 gets its own encoder again from its core bundle or SwarmUI’s automatic download. Or download the Qwen Image 2.1 Core Bundle again: only the 9.35 GB text encoder (shown as 8.71 GB in the downloader) is downloaded, the model and VAE are kept. The Qwen presets do not download models themselves, so they need the renamed or re-downloaded file.</p><h3>NEW AND UPDATED PROMPT GUIDES for ChatGPT, Claude and Gemini</h3><ul><li>NEW — Fibo_2.0_Enchance_Prompt_Feed_For_LLMs.txt: give it to an LLM with your idea (attach the image for an edit) and it writes a Fibo 2.0 JSON caption in Bria’s format: objects, background, lighting, camera, style and every visible text in text_render, for text to image, single-image edits, edits with 2 to 4 images (&lt;image_1&gt;, &lt;image_2&gt;) and masked edits.</li><li>UPDATED — MOSS_Image_Enchance_Prompt_Feed_For_LLMs.txt: a new image edit part: one short instruction that says the change first and then what stays, images named “image 1” and “image 2”, and an empty negative prompt.</li><li>UPDATED — MiniMax_H3_Enchance_Prompt_Feed_For_LLMs.txt: a new FastH3 Omni section. Prompt FastH3 Omni exactly like MiniMax H3 Ref2VA; the presets also take an init image and a locked init audio, and steps or samplers never go into the prompt.</li></ul><h3>FIXES FROM OUR V144 FEATURE TESTS</h3><p>We ran all ten new presets on the ComfyUI our installer gives you today. These problems were found and fixed before this release:</p><ul><li>Fibo 2.0 Dev text to image on the newest ComfyUI: ComfyUI now puts Fibo 2.0’s negative prompt into Bria’s chat template itself, and our node did the same, so the negative was wrapped twice. Our node no longer wraps it. On the newest ComfyUI the four Fibo 2.0 presets now give the same pixels, seed for seed, as before this ComfyUI change.</li><li>FastH3 Omni optional face refine: the face-pass instruction asked for “beard strands”, which gave a woman a beard in our test; the words are gone from the presets and from the node’s default. With FastH3 the old constant 0.55 pass was rejected by our hallucination guard on small faces and changed nothing, so the FastH3 presets now use 0.35, which repaired them (the image preset also uses FastH3’s 8 steps, Euler / simple, for the face pass).</li><li>FastH3 Omni model download links: the model loader pointed ComfyUI’s missing-model download at MiniMax H3 Ref2VA and the text encoder loader at the NVFP4 encoder; they now point to the FastH3 Omni INT8 model and the INT8 Qwen3-VL 32B encoder the presets load.</li><li>Token meter: in presets with RefMods, the meter under the prompt ignored the gallery’s references and showed “text to video” for a reference run; it now counts them.</li><li>Long Lip Sync without a soundtrack now stops with a clear message instead of a Python error.</li><li>Notes and titles in the FastH3 presets now match their settings (speed mode, face refine, seed).</li></ul><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V144.zip from this post.</li><li>Extract it over your existing installer folder and overwrite the older files.</li><li>Close ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat. It updates ComfyUI to the newest version with Fibo 2.0 and FastH3 Omni support, and our ComfyUI-TeaCache and FoleyExtension, plus ComfyUI-Manager. The Fibo 2.0 and FastH3 Omni presets do not run on older ComfyUI.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install/update bundle 100. It updates SECoursesAudioTools with the Fibo 2.0 and MOSS-Image Edit nodes and ComfyUI-KJNodes, and keeps SwarmUI ExtraNodes and ComfyUI-Easy-Use, which the new presets also use, up to date.</li><li>Restart ComfyUI and refresh the browser tab.</li></ol><p>Use our latest SwarmUI Model Downloader (SwarmUI_Model_Downloader_v184.zip or newer) and download the bundles from the SwarmUI Bundles tab (the downloader shows every size in binary GB; the bundle totals below are its numbers, the file sizes are decimal GB):</p><ul><li>FastH3 Omni Core Bundle (first card, 49.78 GB, 9 files): the FastH3 Omni 8-step INT8 ConvRot model (22.2 GB), the INT8 Qwen3-VL 32B text and vision encoder (27.1 GB), the MiniMax H3 video and audio VAEs and the face detection, recognition and restoration models of the optional face refine and mouth pass. BF16, FP8 and NVFP4 versions of the model and the NVFP4 and BF16 encoders are optional downloads. If you downloaded the MiniMax H3 bundle with V183, its INT8 video VAE was Kijai’s file (3.17 GB) saved under the name of Comfy-Org’s file that V184 expects, so the downloader downloads Comfy-Org’s file once (2.81 GB, shown as 2.62 GB).</li><li>Fibo 2.0 Core Bundle (second card, 20.50 GB, 4 files): the Dev and Edit Turbo INT8 ConvRot models (8.4 GB each), the INT8 Qwen3-VL 4B encoder (4.9 GB) and the FLUX 2 VAE. The BF16 files and Comfy-Org’s turbo LoRA for Dev are optional downloads.</li><li>MOSS-Image Core Bundle (third card, 21.91 GB, 7 files): it now also has MOSS_Image_Edit_LoRA.safetensors (0.30 GB), ChenkinLab’s edit LoRA, which the MOSS Image Edit presets use.</li></ul><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>18 October 2026 Update V143</h3><ul><li>17 October 2026 Massively Detailed Windows Tutorial : <a href="https://youtu.be/yOj9PYq3XYM">https://youtu.be/yOj9PYq3XYM</a></li></ul><p>Presets Folder Content In Latest Zip File</p><p>This is a MOSS-Image update. V143 adds six new presets for MOSS-Image, the new open text-to-image model from OpenMOSS: Text To Image with the official 50-step quality recipe, Fast 25 Steps, Turbo 8 Steps, Layout Boxes, and Base model versions of Text To Image and Layout Boxes. It also adds a MOSS-Image prompt guide for ChatGPT, Claude and Gemini, a ComfyUI notebook for Google Colab, an updated AvatarForever preset with a live preview panel, fixes for the VRAM-optimized launcher and the installer, the newest ComfyUI with native MOSS-Image support, and the matching custom-node and SwarmUI extension updates.</p><p>Four native 1920×1088 images from the new Advanced Text To Image preset, with no upscaling: 50 steps, 32 to 41 seconds each on an RTX 5090. Every label and poster line came out exactly as written, and the chalkboard is spelled correctly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wlma78cyYkvKPW4W.jpeg" /></figure><p>Every result in this post comes from a real run of the new presets with their default settings; only the prompt (for the layout, also the boxes), seed and output size were changed. The timings were measured on our RTX 5090 after the models were loaded.</p><p>Here is the complete Presets folder of V143: 96 ready-to-use ComfyUI presets, plus 3 API presets in the api_paid folder. The six new MOSS-Image presets are marked NEW and the updated AvatarForever preset is marked UPDATED:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*p0Uic84i1WoDCZVH.png" /></figure><p>The presets cover image, video, lip sync, audio and music models: AvatarForever, ERNIE Image, FLUX.1 Dev, Kontext and SRPO, FLUX 2, FLUX 3 Klein, Ideogram 4, Krea 2, LTX 2, LTX 2.3, LTX 2.5, MiniMax H3, MiniMax Music 3, MOSS-Image, Phantom Wan, Qwen Image, SDXL, Stable Video Infinity 2 Pro, Wan 2.1, Wan 2.2 and Z-Image.</p><h3>NEW PRESETS — MOSS-Image: Text To Image, Fast, Turbo, Layout Boxes and Base</h3><ul><li><em>“MOSS Image — Advanced Text To Image — 261017”</em></li><li><em>“MOSS Image — Fast 25 Steps — 261017”</em></li><li><em>“MOSS Image — Turbo 8 Steps — 261017”</em></li><li><em>“MOSS Image — Advanced Layout Boxes — 261017”</em></li><li><em>“MOSS Image Base — Advanced Text To Image — 261018”</em></li><li><em>“MOSS Image Base — Advanced Layout Boxes — 261018”</em></li></ul><p>Here are the new and updated presets, with what each one does:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IkOu9J-QRL6r9OVw.png" /></figure><p>MOSS-Image is a new open image model from OpenMOSS: a 5B diffusion transformer that reads your prompt with the Qwen3.5–9B language model and uses the FLUX 2 VAE. It renders clear, accurate text inside images, accepts long and detailed prompts of up to about 4,000 tokens, and, as an experimental feature of the model, can place objects, characters and text in boxes you draw. It makes images from text; it does not edit images or use reference images.</p><p>SFT and Base: SFT is the official fine-tune of Base on curated data, and the four main presets use it; in the official comparison it groups text more clearly. Base is the raw model without post-training, with more varied images, meant for LoRA training; two Base presets are included.</p><p>All six presets share the same easy controls:</p><ul><li>One prompt box: plain language or tags, and in the Text To Image presets also a JSON layout. Long, detailed prompts work best. Put exact text for signs, labels and posters in quotes, and use real line breaks to control the lines.</li><li>Output size: choose an aspect ratio and megapixels, or type the width and height. Any aspect ratio works; 1 to 2 megapixels is recommended. Sizes snap to multiples of 16.</li><li>Batches: + Prompt adds more prompts to one run, a folder of .txt files runs a whole prompt list (JSON layouts too), and Images per prompt makes several images of each prompt.</li><li>Prompt syntax: random choices {red|blue} follow the seed in the main prompt, added prompts and folder prompts, so a fixed seed repeats the same prompt. Schedules [red:blue:0.5] and alternation [red|blue] work in plain-text prompts of the Text To Image presets. Weights like (word:1.2) are removed because MOSS-Image’s text encoder would read them as words; describe emphasis in words instead.</li><li>JSON layouts in the Text To Image presets: a complete MOSS-Image layout is sent unchanged, only without extra spaces. Pixel-based layout drafts from the official MOSS prompt enhancer or our prompt guide are converted to the model’s format automatically and scaled to your output size. In the Layout Boxes presets, the prompt box holds only the caption of the whole image.</li><li>Automatic SFT de-band: the SFT model can draw faint vertical color stripes on skies, fog and plain walls. The presets apply a community fix to SFT automatically and skip it for Base.</li><li>Seed: fixed, or randomized on every run, with an optional variation seed and strength.</li><li>Optional LoRA stack for MOSS-Image LoRAs, off by default.</li><li>Attention set to PyTorch: it matched the official MOSS-Image images in our tests, and it also prevents black images with the BF16 files when ComfyUI starts with Sage attention, as our launchers do (RTX 3090 test). Keep it on PyTorch in your own BF16 workflows.</li><li>The model, text encoder, VAE and latent are already connected. To use the BF16 files, switch the model loader. Use Base only in the 50-step presets: the Fast and Turbo recipes were made for SFT.</li></ul><p>Here is the Advanced Text To Image preset after a real run of a magazine cover prompt:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*b51GdQgtCTiliA5-.png" /></figure><p>All five text lines came out exactly as written, from the huge WANDER title to the small issue line. The 1024×1536 image took 21.8 seconds with the official 50-step recipe.</p><p>Fast 25 Steps, Turbo 8 Steps and Base: here is the same prompt and seed through all four text-to-image presets at 1920×1088:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-CjBqENHNmaMhDLG.jpeg" /></figure><ul><li>Advanced Text To Image (SFT, 50 steps, Euler / normal, CFG 3, shift 5, the official recipe): the most faithful results, especially for text and designs. 34.3 seconds here, 12.9 seconds at 1024×1024.</li><li>Fast 25 Steps (ER-SDE / simple, CFG 3, Temporal Score Rescaling 0.95, layer skip 22 and 24 from 10% to 70% of the steps): 16.6 seconds here, about half the time of the 50-step preset. Large text holds up well; very small text can occasionally lose a letter.</li><li>Turbo 8 Steps (the community 8-step LoRA by lodestones at 0.7, Euler / simple, CFG 1): 3.4 seconds here, about a tenth of the 50-step time. The 8-step LoRA is a work in progress, so its images differ more from the 50-step ones and can show extra signs.</li><li>Base Advanced Text To Image (Base model, the official 50-step recipe): more varied images; 32.5 seconds here. For text-heavy designs, start with SFT: in the official Base versus SFT example, SFT groups text more clearly.</li></ul><p>Layout Boxes: draw a box for each object, character or line of text and describe it. The main prompt becomes the caption of the whole image:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CTiCkgceqmXLjjsx.png" /></figure><ul><li>Each box is an object or exact text, with its own description. To turn object boxes into characters, type their numbers in character_boxes, here 8 for the barista.</li><li>The canvas follows the output size, and the preset builds MOSS-Image’s layout JSON from your boxes.</li><li>The box editor needs Settings → Modern Node Design (Nodes 2.0). Turn it on once.</li><li>Boxes guide placement. They are not masks, so positions can vary a little.</li></ul><p>Here is the result of that layout in the same preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ufqYj4LStjVIO24I.png" /></figure><p>Every box landed where it was drawn: the headline, three drinks with their exact names and prices, and the waving barista as a character. The 1536×1024 image took 22.6 seconds.</p><h3>UPDATED PRESET — AvatarForever: Live Preview Panel and Tested Mouth-Movement Settings</h3><p><em>“AvatarForever Unified Image or Text + Audio To Long Video — 260920”</em></p><ul><li>New Live Preview &amp; Final Video panel to the right of the generation controls. It shows a frame preview after each sampling step, starting with the first, then plays the finished MP4.</li><li>New tested preset settings: a 512×768 canvas and 16 latent frames per chunk, instead of 832×1248 and 4. In our singing demo, 4-frame chunks kept the lips closed, while 16-frame chunks gave sustained mouth movement through the whole 18-second song. At 1024×1024 the mouth stayed mostly closed even with 16-frame chunks, so check a short sample when you change the image, audio, resolution or chunk size. Larger chunks need more memory; 4-frame chunks still work for speech and use less.</li><li>Sharper previews: the live preview uses the small taeltx2_3 preview decoder when it is installed. It is now in the LTX 2.3 Avatar Forever Core Bundle of our SwarmUI Model Downloader V183. Without it, the preview uses a coarser latent view.</li><li>Your own AvatarForever workflows keep working and now show live previews on the sampler. Set live_preview to off to turn them off.</li></ul><h3>NEW PROMPT GUIDE — MOSS-Image Feed for ChatGPT, Claude and Gemini</h3><p>A new MOSS_Image_Enchance_Prompt_Feed_For_LLMs.txt is included. Give it to ChatGPT, Claude or Gemini together with your request, and it writes prompts that MOSS-Image understands:</p><ul><li>Detailed captions for photos, illustrations and designs, with the exact visible text in quotes and where and how it appears.</li><li>Layouts as pixel-based JSON drafts for the Text To Image presets, which convert them to MOSS-Image’s own format automatically, or a box-by-box plan to draw in the Layout Boxes editor.</li><li>Tag-style prompts for anime, in the recommended tag order.</li><li>The right settings: output size, CFG for short prompts, the negative prompt and when to use SFT or Base.</li></ul><h3>NEW — ComfyUI Colab Master Notebook for Google Colab</h3><p><em>“ComfyUI_Colab_Master_Notebook.ipynb”</em></p><p>A new Google Colab notebook runs ComfyUI with ComfyUI-Manager on a Colab GPU and opens it in your browser through a Cloudflare link. It sets up Qwen-Image 2.1 for you:</p><ul><li>Install: ComfyUI (the newest stable release, or master if you choose it) with ComfyUI-Manager and cloudflared, which makes the link without an account. Colab’s own PyTorch is kept.</li><li>Qwen-Image 2.1 in the right size: the official ComfyUI files are picked from your GPU memory and system RAM: INT8 with the W4A8 text encoder (14.2 GB) when either is below 20 GiB, as on a T4; INT8 with the INT8 text encoder (17.3 GB), the official template’s files, when both have at least 20 GiB; or full BF16 (32.4 GB) when you choose it. Broken downloads resume when you run the cell again.</li><li>Any other model: paste up to four direct download links, such as LoRAs or upscalers, and choose the folder or let the notebook read it from the link.</li><li>Start: you get a <a href="https://trycloudflare.com/">trycloudflare.com</a> link, and ComfyUI opens with the Qwen-Image 2.1 text-to-image workflow already loaded, with the official settings (Euler, simple, 25 steps, CFG 1, 1024×1024).</li><li>Keep your work: one cell packs your outputs and saved workflows into a ZIP to download. The notebook also explains how to upload images and videos, and what is kept when the runtime ends.</li></ul><p>Open Google Colab, choose Upload notebook, select the file, pick a GPU runtime (T4 or better) and run steps 1 to 4 in order. T4 GPUs can be allocated on free accounts, but availability is not guaranteed, and Colab says free runtimes used mainly through an external web UI may be terminated. The Cloudflare link is public, so don’t share it.</p><h3>LAUNCHER AND INSTALLER FIXES</h3><ul><li>Windows_Run_VRAM_Optimized.bat: option 3 (normal VRAM), and VRAM choice 3 in Custom mode, passed — normalvram, which ComfyUI removed in version 0.22.0, so ComfyUI stopped at once with “unrecognized arguments: — normalvram”. Both now start ComfyUI with its default memory management and no extra flag.</li><li>Windows_Run_VRAM_Optimized.bat also describes the memory modes as current ComfyUI uses them. LOWVRAM’s — lowvram has no effect while dynamic VRAM is on, the default on NVIDIA GPUs, and otherwise only runs the text encoders on the CPU; NOVRAM turns dynamic VRAM off. Pressing Enter in the cache menu now keeps ComfyUI’s own default, RAM-pressure caching, instead of the older classic caching; choose 1 for classic caching.</li><li>Windows_Install_Or_Update_ComfyUI.bat: the pip update and the uv install now ignore your personal pip settings (the pip.ini in your user profile and PIP_ environment variables), so a custom package index set there for another project can no longer stop the installer at this step. Its Windows requirements tutorial link now points to the current video: <a href="https://youtu.be/ptuLL0C72Ms">https://youtu.be/ptuLL0C72Ms</a></li></ul><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V143.zip from this post.</li><li>Extract it over your existing installer folder and overwrite the older files.</li><li>Close ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat. It updates ComfyUI to the newest version, with MOSS-Image support. The MOSS-Image presets do not run on older ComfyUI.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install/update bundle 100. It updates SECoursesAudioTools with the new MOSS-Image nodes and the AvatarForever Live Preview &amp; Final Video node that the updated AvatarForever preset uses, and keeps FoleyExtension, SwarmUI ExtraNodes and ComfyUI-Easy-Use, which the MOSS-Image presets also use, up to date.</li><li>Restart ComfyUI and refresh the browser tab.</li><li>For the Layout Boxes editor, turn on Settings → Modern Node Design (Nodes 2.0).</li></ol><p>Use our latest SwarmUI Model Downloader (SwarmUI_Model_Downloader_v183.zip or newer) and download the MOSS-Image Core Bundle, the first card of the SwarmUI Bundles tab. It has the six files the six presets use, about 23.2 GB in total (the downloader counts in GiB, so it shows 21.64 GB):</p><ul><li>moss_image_sft_int8_convrot.safetensors (5.41 GB), the SFT model in INT8 ConvRot.</li><li>moss_image_base_int8_convrot.safetensors (5.41 GB), the Base model in INT8 ConvRot.</li><li>qwen3.5_9b_int8_convrot.safetensors (9.75 GB), the Qwen3.5–9B text encoder in INT8 ConvRot.</li><li>Flux/flux2-vae.safetensors (0.34 GB), the FLUX 2 VAE, the same file the FLUX 2 and FLUX 3 presets use.</li><li>MOSS_Image_SFT_DeBand_L37–41.safetensors (1.13 GB), the SFT de-band fix.</li><li>MOSS_Image_SFT_8_Steps_DMD_LoRA.safetensors (1.19 GB), the 8-step LoRA of the Turbo preset.</li></ul><p>The INT8 ConvRot model files are about half the size of BF16, and in our RTX 3090 test they were about twice as fast (48 versus 99 seconds at 50 steps). The BF16 models, a smaller W4A8 text encoder (select it in the text encoder loader), the small-decoder FLUX 2 VAE and the extra de-band fix for layers 32–36 are optional downloads under Image Generation Models → MOSS-Image Models. The de-band author recommends the extra fix for images 1536 pixels wide or larger and for big, smooth skies; select it in the de-band node.</p><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>17 October 2026 Update V142</h3><p>This is a FLUX 3 Klein update. V141 adds three new presets for Black Forest Labs’ new FLUX 3 Image Klein model: Text To Image, Image Edit with up to 10 reference images, and Layout Boxes. It also adds a FLUX 3 Klein prompt guide for ChatGPT, Claude and Gemini, our new HQ INT8 ConvRot FLUX 3 Klein model, ComfyUI 0.41.0 with native FLUX 3 support, and the matching custom-node and SwarmUI extension updates.</p><p>A native 3840×2160 image from the new Text To Image preset, with no upscaling: 16 steps in 23 seconds on an RTX 5090.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LPWiWNXFLnBHQDm9.jpeg" /></figure><p>Every result in this post comes from a real run of the new presets with their default 16 steps, Euler / simple, CFG 1. The timings were measured on our RTX 5090 after the models were loaded.</p><h3>NEW PRESETS — FLUX 3 Klein: Text To Image, Image Edit and Layout Boxes</h3><p><em>“FLUX 3 Klein — Advanced Text To Image — 261015”</em></p><p><em>“FLUX 3 Klein — Advanced Image Edit — 261015”</em></p><p><em>“FLUX 3 Klein — Advanced Layout Boxes — 261015”</em></p><p>FLUX 3 Image Klein is a 5.6B image model with the Qwen3-VL-4B text encoder. One model does text to image, image editing, up to 10 reference images, layout boxes and text rendering. Our presets run the fast distilled model with its recommended 16 steps and no guidance (CFG 1), using Euler / simple.</p><p>All three presets share the same easy controls:</p><ul><li>One prompt box with an image gallery: write the whole scene. Add up to 10 images and refer to them as @image1 … @image10 (@img1, @picture1 and @pic1 also work). With no image, it is text to image.</li><li>Output size: choose an aspect ratio and megapixels, or type the width and height. Sizes snap to multiples of 16.</li><li>Batches: + Prompt adds more prompts to one run, a folder of .txt files runs a whole prompt list, and Images per prompt makes several images of each prompt.</li><li>Advanced prompt syntax: weights (ceramic:1.1), random choices {red|blue}, schedules [red:blue:0.5], alternation [red|blue] and &lt;break&gt;. Random choices follow the seed, so a fixed seed repeats the same prompt.</li><li>Seed: fixed, or randomized on every run, with an optional variation seed and strength.</li><li>Reference size: one Reference megapixels setting resizes every reference image and keeps its shape (1 MP by default, up to 4 MP). It is separate from the output size.</li><li>Optional LoRA stack for FLUX 3 LoRAs, off by default.</li><li>The model, text encoder, VAE and latent are already connected. There is nothing else to set up.</li></ul><p>Here is the Text To Image preset after a real run of a poster prompt with English and Turkish text:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hnKeQjjD4faF9NtE.png" /></figure><p>All four text phrases came out exactly as written, including the Turkish characters. The 1024×1536 image took 3.2 seconds. The first run takes longer because it loads the models.</p><p>Image Edit: the same workflow with your image in the gallery. Here it turns a celadon museum vase cobalt blue and keeps its shape, handles, framing and shadow:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SfcLsi1E-S4DzRgV.png" /></figure><ul><li>Say what changes and what stays, and refer to your photo as @image1.</li><li>Match the output aspect ratio to your photo to keep its framing. This 1024×1536 edit took 5.2 seconds.</li><li>The preset opens with a sample card. Add your own image with + Add reference images and remove any card marked “file missing”.</li></ul><p>Up to 10 reference images in one prompt: here ten separate museum photos became one display wall, each object in its requested niche:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sfCfiRI7UqlSENmP.png" /></figure><p>All ten objects appear in the requested order and closely match their photos. The 1680×944 image took 41 seconds with all ten references at 1 MP each. The reference photos are public domain (CC0) images from The Cleveland Museum of Art.</p><p>Layout Boxes: draw a box for each object or line of text, describe it, and refer to it in the caption as &lt;obj_1&gt;, &lt;obj_2&gt; and so on:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qbliLIKQxHVUDBpT.png" /></figure><ul><li>Each box is an object or exact text, with its own description. The preset builds FLUX 3’s layout table from your boxes, and the boxes follow the output size.</li><li>The box editor needs Settings → Modern Node Design (Nodes 2.0). Turn it on once.</li><li>Boxes guide placement. They are not masks, so positions can vary a little.</li></ul><p>Here is the result of that layout in the same preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WxDhrIo4jCkWDBvD.png" /></figure><p>All four terrariums and captions landed where their boxes are, and the four captions are spelled exactly. The 1024×1024 image took 2.5 seconds.</p><h3>NEW MODEL — Our FLUX 3 Klein HQ INT8 ConvRot</h3><p><em>“FLUX-3-Image-Klein-INT8-ConvRot-HQ.safetensors”</em></p><p>The presets use our own INT8 ConvRot conversion of the distilled FLUX 3 Klein model. On the updated ComfyUI it loads with the normal Load Diffusion Model node, so no special loader is needed.</p><ul><li>6.07 GB instead of 11.1 GB for the original BF16 model. While denoising at up to 1024×1024, the model alone peaked at about 6 GiB of VRAM instead of about 10.6 GiB for BF16 (RTX 3090 test).</li><li>About 2× faster per step than BF16 (1.9–2.3× in our RTX 3090 tests, model only).</li><li>Closer to BF16 than the official INT8 file: in our test of 40 sampling states, its mean output error against BF16 was 2.14%, versus 3.05% for the official INT8 ConvRot file with its default settings and 2.27% with matched attention. One held-out editing case favored the official file with matched attention. In exchange, our file is 0.25 GB larger and was 6–22% slower per step in the same RTX 3090 test. These are numerical measurements made with the BF16 text encoder, not image-quality percentages.</li><li>The first text block and the last joint block, the most sensitive in our tests, stay in BF16, as do the small conditioning, normalization and output layers. The other 164 linear layers use native INT8 ConvRot.</li></ul><h3>NEW PROMPT GUIDE — FLUX 3 Klein Feed for ChatGPT, Claude and Gemini</h3><p>A new FLUX_3_Klein_Enchance_Prompt_Feed_For_LLMs.txt is included. Give it to ChatGPT, Claude or Gemini together with your request and images, and it writes prompts that FLUX 3 Klein understands:</p><ul><li>Complete scene captions for text to image, with the most important details first.</li><li>Edit prompts that say what changes and what must stay, and multi-reference prompts that give every image a clear role.</li><li>Exact text for posters and designs, in the original language and spelling.</li><li>Layout prompts in the FLUX 3 box format, with the &lt;obj_1&gt; labels these presets use.</li><li>The right settings and the ComfyUI / SwarmUI details, such as @image1 meaning FLUX 3’s first reference image.</li></ul><h3>AvatarForever Preset Now Has Live Preview Feature</h3><ul><li>2:3 and 3:2 aspect ratios works best — 1:1 not that good</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Mbkrv9iJuSfc-Mhr.png" /></figure><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V141.zip from this post.</li><li>Extract it over your existing installer folder and overwrite the older files.</li><li>Close ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat. It updates ComfyUI to the newest version, with FLUX 3 support. The FLUX 3 presets do not run on older ComfyUI.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install/update bundle 100. It updates SECoursesAudioTools with the new FLUX 3 nodes and keeps FoleyExtension, SwarmUI ExtraNodes and ComfyUI-Easy-Use, which the presets also use, up to date.</li><li>Restart ComfyUI and refresh the browser tab.</li><li>For the Layout Boxes editor, turn on Settings → Modern Node Design (Nodes 2.0).</li></ol><p>Use our latest SwarmUI Model Downloader (SwarmUI_Model_Downloader_v181.zip or newer) and download the FLUX 3 Klein Core Bundle, the first card of the SwarmUI Bundles tab. It has the three files all three presets use, about 11.7 GB in total (the downloader counts in GiB, so it shows 10.85 GB):</p><ul><li>FLUX-3-Image-Klein-INT8-ConvRot-HQ.safetensors (6.07 GB), our HQ INT8 ConvRot model.</li><li>qwen3vl_4b.safetensors (5.24 GB), the FP8 Qwen3-VL-4B text encoder.</li><li>Flux/flux2-vae.safetensors (0.34 GB), the FLUX 2 VAE, which FLUX 3 uses too.</li></ul><p>The text encoder is the same file Krea 2 uses and the VAE is the FLUX 2 VAE, so if you already have those, only the 6.07 GB model is new.</p><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>10 October 2026 Update V140</h3><p>ComfyUI V140 is ready. Here is everything we changed since V138: 4 new MiniMax H3 presets, optional MiniMax H3 RefMods (reusable reference files) in our MiniMax H3 presets, a native sparse attention engine in our MiniMax H3 Speed Optimizer, fixes that keep our nodes working on the newest ComfyUI (tested on ComfyUI 0.39.0), and fixes for every preset that failed when we ran all presets with real generations.</p><h3>New MiniMax H3 Presets</h3><p>MiniMax H3 Lip Synch — Turbo 4 Steps (672x1216 and 832x1248)</p><p>Here is the 672x1216 preset after a real run with its demo audio and portrait:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RI1iV-lwd-SsbKLW.png" /></figure><ul><li>Fast talking videos from your audio: upload an audio file, optionally add an image, describe the scene and run. The video length follows your audio. Without an image it works as audio + text. With a video in the reference gallery, it makes a new performance from that video.</li><li>Turbo 4 steps by default with the LightX2V FL2VA Turbo 4-step v1.2 LoRA, at 24 fps. A 15-second video took about 165 seconds at 672x1216 on an RTX 5090. Set “4. QUALITY / SPEED / VRAM” to FULL QUALITY for the 50-step path.</li><li>Two portrait resolutions: 672x1216, or 832x1248 for more detail. These are 4-step portrait versions of our existing MiniMax H3 Lip Synch preset (1024x1024, 8 steps).</li></ul><p>The 832x1248 preset saves the original video first, then runs the new CodeFormer mouth pass:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LVW45SHSV2JBLpta.png" /></figure><ul><li>The 832x1248 preset adds a CodeFormer mouth pass (on by default). It restores only the mouth area of the finished video and saves it as a second MP4, next to the original video.</li><li>Also included: our optional face refinement (off by default) and the new optional RefMods.</li><li>Tip: for your own audio, put its exact words in the &lt;d&gt;[English] … &lt;/d&gt; part of the scene prompt, or clear the scene prompt to use the automatic one. 5–15 seconds per generation is recommended; longer audio still works in one pass.</li></ul><h3>MiniMax H3 Long Lip Synch — Native Latent Continuation — Full Audio</h3><p>Here is the Long Lip Synch preset with its default settings:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Eqd8QSQa6jZLCHiF.png" /></figure><ul><li>One full soundtrack in, one continuous video out: give it a complete audio file plus a portrait image or gallery references. It makes one continuous 24 fps lip-synced video as long as your audio. There is no duration setting and no cutting.</li><li>How it stays continuous: it generates overlapping windows of about 10 seconds (243 frames) and carries the last 1.6 seconds (39 frames) of each window forward as clean latents, so nothing is re-encoded between windows. The audio is encoded once, and your original soundtrack is put back into the final MP4.</li><li>Works with references and RefMods in every window, unlike the TaoMate streaming preset. The first-frame image anchors only the start.</li><li>Defaults: 832x1248 and Turbo 4 steps (FL2VA 4-step v1.2, or Ref2VA 4-step v0.1 when you use references). Use 512x768 for faster drafts. SOL sparse attention and FirstBlockCache are separate switches, both off by default.</li><li>Outputs: the final MP4 with your original audio, plus a JSON report with the timing of every window. On an RTX 5090, 30 seconds of 512x768 video took 105 seconds with normal attention.</li><li>Experimental: very long videos can change framing over time. In our 15-clip test, 13 clips passed; the 45 and 60-second clips at 832x1248 drifted into close-ups.</li></ul><h3>MiniMax H3 RefMod Builder — Images, Video, Audio</h3><p>Add your source media, name your RefMod and run. Only the MiniMax H3 VAEs are used:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ckUpk5YYD9OYl-1z.png" /></figure><ul><li>Build your own RefMod in one run: add photos of one subject, a video (its soundtrack is included), audio, or a mix. Name it and run. Only the MiniMax H3 VAEs run: no diffusion model, no text encoder and no training.</li><li>Options: source resolution (512 by default), “one character” (stacks the photos into one reference) or “separate references” (each photo gets its own picture label), and optional compression (32, 16 or 8) for fewer tokens at the cost of detail.</li><li>Ready to use: files are saved to ComfyUI/models/refmods and existing files are never overwritten. Pick the file in any MiniMax H3 preset’s Optional RefMods node and write @refmod1 in your prompt.</li></ul><h3>MiniMax H3 RefMods: Reusable References</h3><p>A RefMod is a small .safetensors file that stores already-encoded MiniMax H3 references: images, video and/or audio of a character, scene or voice. Build it once, then reuse or share it in any MiniMax H3 preset without encoding the source media again. It is not a LoRA and needs no training.</p><p>Here is the new Optional RefMods node with an imported RefMod, used as @refmod1 in the prompt:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ukcHJKa1d0h3raEU.png" /></figure><ul><li>New “MiniMax H3 Optional RefMods” node in all 9 existing MiniMax H3 generation presets (Text To Video, Image or Video To Video, References To Video, Audio Only, Text With References To Image and Lip Synch) and in the new presets. It is empty by default, so your results do not change unless you add a RefMod.</li><li>Easy picker: “+ Add RefMod”, “Import files” and “Refresh” buttons. Each row has an on/off box, the file (with its contents, for example “image + audio”), a strength from 0 to 1, a visual/audio filter and up/down ordering. Imported files go to ComfyUI/models/refmods and are checked before use.</li><li>Use them in your prompt with @refmod1, @refmod2 and so on. They are numbered after your gallery images, videos and audio.</li><li>Strength sets how much reference detail is kept. It is not a LoRA weight.</li><li>RefMods alone are enough: the reference mode and auto mode nodes count RefMods as references, so a RefMod without any gallery image switches generation to the references (Ref2VA) path.</li><li>Compatible files: RefMods made with ComfyUI-MiniMaxH3Mod and the Fantastic MiniMaxH3 PromptBuilder load too, without installing those node packs.</li><li>TaoMate streaming does not use RefMods; its preset note now points you to the new Long Lip Synch preset for RefMods with long audio.</li></ul><h3>MiniMax H3 Speed Optimizer: Native Sparse Attention</h3><p>Here is the MiniMax H3 Speed Optimizer node, shown in the Text With References To Image preset, where it is off by default:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rOhzdVL2BsZqkngw.png" /></figure><ul><li>SOL sparse attention now runs on ComfyUI’s own MiniMax H3 sparse attention (our ComfyUI-TeaCache fork). The old bundled attention kernels are gone, so there is nothing extra to compile.</li><li>Faster than our old SOL: the old version repeated work and was slower than normal attention at 832x1248 (207 s against 116 s). With the new default protected SOL settings, a 10-second 832x1248 Turbo 4-step video took 99 seconds instead of 111 seconds with normal attention (1.12x, RTX 5090, FirstBlockCache off for this test). At 512x768 the gain was 1.06x. Sparse attention and caching are approximations, so check the quality of your own results.</li><li>Auto mode measures before it switches: it uses sparse attention only when it is at least 5% faster. The Text To Video and Image or Video To Video presets have the optimizer on in auto mode, so they now use the native path automatically when it pays off.</li><li>FirstBlockCache works together with other speed-ups: it now keeps patches that are already on the model, such as the sparse attention added by the lip-sync Turbo node, instead of replacing them.</li><li>Old presets load unchanged. The main switch is now called “H3 optimizations” instead of “4x speedup”, and “native” is the new default backend.</li></ul><h3>Other MiniMax H3 Improvements</h3><ul><li>Correct colours in streamed MiniMax H3 videos: TaoMate streaming and long lip-sync videos are now tagged with the colour format they are encoded in (BT.601 limited range), so players read the colours correctly.</li><li>No torchaudio needed for MiniMax H3 audio: init audio, streaming and long lip sync now resample any soundtrack with ComfyUI’s own resampler, with bit-identical results (details below).</li></ul><h3>Works With the Latest ComfyUI</h3><p>ComfyUI removed torchaudio and changed its attention and Linear layer code. We updated our nodes so everything keeps working on the newest ComfyUI:</p><ul><li>No more torchaudio needed: our MiniMax H3 streaming and long lip-sync nodes (SECoursesAudioTools), the init audio nodes (FoleyExtension) and the audio models in our ComfyUI-WanVideoWrapper fork (HuMo, Ovi, FantasyTalking, MultiTalk and LongCat) now resample audio with ComfyUI’s own resampler. Before, on installs without torchaudio, they failed on any soundtrack that was not already at the model’s sample rate, and HuMo and Ovi did not load at all. The output is bit-identical: 88 test comparisons had zero difference, and a real MiniMax H3 Turbo 4 video was pixel-identical on all 120 frames.</li></ul><p>Here are the new torchaudio-free Mel-Band RoFormer nodes inside the LTX2.3 audio lip-sync preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VWAC4R6MzR5xsF4S.png" /></figure><ul><li>New torchaudio-free Mel-Band RoFormer vocal separation nodes (SECoursesAudioTools: “SE Mel-Band RoFormer Model Loader” and “SE Mel-Band RoFormer Sampler”). They load the same checkpoint and give bit-identical vocals, so the LTX2 and LTX2.3 audio lip-sync presets keep their vocal separation even where the original Mel-Band RoFormer node pack cannot load.</li><li>GGUF models work again (our ComfyUI-GGUF fork): on the latest ComfyUI every quantized GGUF model failed with “unexpected keyword argument ‘input_act’”. Fixed and verified bit-exact on 57 test cases.</li><li>FLUX 2 Mistral GGUF text encoder loads again (ComfyUI-GGUF): the GGUF Tekken tokenizer now gives ComfyUI the special token count it needs. Token IDs match the official tokenizer.</li><li>Much faster FLUX 2 GGUF on limited VRAM (ComfyUI-GGUF): GGUF text encoders are now fully unloaded when VRAM is needed, instead of leaving 6.6 GB in VRAM. FLUX 2 Low RAM sampling went from 200–309 s/it to 1.75 s/it on a 32 GB GPU.</li><li>LTX 2.3 Foley presets fixed: ComfyUI-LTXVideo is now installed from our fork, which fixes the STG error “‘AttentionTensorContainer’ object has no attribute ‘shape’”. Existing installs switch to the fork automatically on their next update.</li><li>ControlNet Aux keeps the MIT albumentations package (our comfyui_controlnet_aux fork): upstream switched to AlbumentationsX, which is AGPL-3.0, sends usage telemetry by default and conflicts with insightface (used by ReActor).</li></ul><h3>Preset Fixes and Updates</h3><p>We ran every preset with real generations on the latest ComfyUI and fixed every one that failed. The FLUX 2 Low RAM presets now free VRAM between prompt encoding and sampling:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RAJ7cg1pY3F8rAGX.png" /></figure><ul><li>FLUX 2–8 Steps Fast HQ / Realism — Low RAM: two new “Free VRAM before sampling” nodes unload the models after the prompts are encoded. Speed went from 200–309 s/it to 1.75 s/it, 35–75 seconds per image on a 32 GB GPU. The Realism Low RAM preset also had the HQ preset’s guidance of 7; it is now 1, like the normal Realism preset.</li><li>Krea 2 Raw Base, Krea 2 Turbo and Krea 2 Turbo Image Edit: the saved sampler values were shifted by one position (for example steps became 1 and CFG became “euler”). Fixed.</li><li>LTX2 Image To Video and LTX2 Text To Video 8 Steps: the seed was empty, so generation failed. Fixed.</li><li>LTX2.3 Extend Any Video With Audio: the input video is now resized to a multiple of 64, so the source and the half-resolution second stage always match (1280x720 is unchanged).</li><li>LTX2.3 Foley Video To Audio: the video loader read the frame limit (169) as “skip first frames”. Fixed.</li><li>LTX2.5 Text To Audio: the text encoder and prompt enhancer were connected to the wrong inputs, and the enhance prompt option held a model file name. Fixed.</li><li>LTX2.5 IC LoRA Pixel Spatial Upscaler 2x (FAST Distilled and Slower HQ): the VAE paths now point to the LTX-2 folder, where our model downloader saves them.</li><li>LTX2 and LTX2.3 Audio Lip Synch Image To Video (3 presets): now use our new torchaudio-free Mel-Band RoFormer nodes for vocal separation.</li><li>AvatarForever Unified Image or Text + Audio To Long Video: now renders an exact 832x1248 portrait by default, instead of a 768x512 pixel budget that followed the input image’s shape.</li></ul><h3>Installer Improvements</h3><ul><li>Smarter node updates (Windows, RunPod and Massed Compute): the updater now asks GitHub for each node’s default branch and fetches only that branch. When we move a node to a different fork, as with ComfyUI-LTXVideo in this release, existing installs switch over cleanly.</li><li>Reviewed fork updates: our forks of rgthree-comfy, ComfyUI-Easy-Use, comfyui_controlnet_aux and RES4LYF are synced with their latest upstream code. We read every new upstream commit for safety first, then re-ran our presets with identical prompts: 85 of 85 passed.</li><li>comfy-kitchen minimum version raised to 0.2.37, which the new MiniMax H3 sparse attention and long lip sync need.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ca62298f4176" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Install ComfyUI and Learn FLUX 3 Klein: The Complete Beginner Tutorial]]></title>
            <link>https://medium.com/@furkangozukara/install-comfyui-and-learn-flux-3-klein-the-complete-beginner-tutorial-cc510e8a14e3?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/cc510e8a14e3</guid>
            <category><![CDATA[tutorial]]></category>
            <category><![CDATA[comfy-ui]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 02:31:50 GMT</pubDate>
            <atom:updated>2026-10-19T02:31:50.181Z</atom:updated>
            <content:encoded><![CDATA[<h3>Full tutorial link &gt; <a href="https://www.youtube.com/watch?v=nxsCBTcCVp4">https://www.youtube.com/watch?v=nxsCBTcCVp4</a></h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FnxsCBTcCVp4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DnxsCBTcCVp4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FnxsCBTcCVp4%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/657596fb4635b2bec1751f4e16ba830f/href">https://medium.com/media/657596fb4635b2bec1751f4e16ba830f/href</a></iframe><p>Install ComfyUI on Windows and make precise FLUX 3 Klein images: 16 core examples, RTX 5090 tests and launcher tuning. The package includes RunPod, SimplePod, Massed Compute and Linux routes.</p><h3>Links:</h3><ul><li>ComfyUI package: [ <a href="https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709">https://www.patreon.com/SECourses/posts/comfyui-auto-2-105023709</a> ]</li><li>Model downloader: [ <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a> ]</li><li>Windows requirements tutorial: [ <a href="https://youtu.be/ptuLL0C72Ms">https://youtu.be/ptuLL0C72Ms</a> ]</li><li>Requirements installer: [ <a href="https://www.patreon.com/SECourses/posts/1-click-windows-172290854">https://www.patreon.com/SECourses/posts/1-click-windows-172290854</a> ]</li><li>Cloud setup tutorials: [ <a href="https://youtu.be/9C_6qNKjgpA">https://youtu.be/9C_6qNKjgpA</a> ]</li></ul><h3>Info</h3><ul><li>You will learn installation, models, captions, image edits, layouts, batches, shared storage, launch arguments and updates. Follow the complete route from the download post to a working image. Open saved workflows, place references, inspect enlarged results and compare changes using sliders and labeled boards. The examples show exact lettering, object preservation and ten-reference composition.</li><li>Main topics include HQ INT8 ConvRot, Gemini, seeds, LoRAs and VRAM. Benchmarks: up to 2.3x on RTX 3090 and 1.82–2.18x on RTX 5090, for warm transformer steps only; loading, text encoding and image decoding excluded. Learn what — reserve-vram changes, why a saved BAT needs a server restart, and how to verify a small workflow afterward. Change one option at a time and compare the same task.</li><li>Use the chapters to jump to each setup step, feature and comparison.</li></ul><h3>Chapters:</h3><ul><li>00:00:00 Precise edits and 16 FLUX 3 examples</li><li>00:02:47 Open a preset and generate your first image</li><li>00:04:00 Package and Windows requirements</li><li>00:04:54 Download and extract the installer</li><li>00:06:26 Install ComfyUI and Python</li><li>00:07:11 Finish Torch and CUDA setup</li><li>00:07:56 Install custom nodes with bundle 100</li><li>00:09:08 Download the model downloader</li><li>00:10:02 Start and explore the downloader</li><li>00:11:19 Set model folders and ComfyUI structure</li><li>00:11:45 Select the FLUX 3 core bundle</li><li>00:12:36 Download and verify three models</li><li>00:13:06 Check model filenames and folders</li><li>00:13:56 Use the URL Downloader</li><li>00:14:27 Start ComfyUI and explore nodes</li><li>00:15:20 Open the advanced image preset</li><li>00:16:03 Describe the aurora and inspect loaders</li><li>00:16:41 Set shape, megapixels and dimensions</li><li>00:17:17 Seed, sampler, steps and guidance</li><li>00:17:46 Generate and inspect the aurora</li><li>00:19:05 Create a crystal kingfisher</li><li>00:19:58 Generate exact poster lettering</li><li>00:21:06 HQ INT8 ConvRot conversion</li><li>00:21:21 RTX 3090 speed and mixed precision</li><li>00:22:34 RTX 5090: BF16 versus HQ INT8</li><li>00:23:44 RTX 5090 live generation</li><li>00:25:08 Give the prompt guide to your LLM</li><li>00:26:18 Use Gemini for captions</li><li>00:27:15 Expand scene, poster and edit ideas</li><li>00:28:30 Generate from the enhanced caption</li><li>00:29:38 Open image editing and load a reference</li><li>00:30:22 Describe changes and preserved details</li><li>00:31:29 Inspect the before-and-after slider</li><li>00:32:16 Remove, move and replace desk objects</li><li>00:34:03 Reinterpret a painting as photography</li><li>00:35:15 Keep cat identity in a new setting</li><li>00:36:32 Combine identity and style references</li><li>00:38:22 Transform a landscape into winter</li><li>00:39:23 Compare geography and seasonal light</li><li>00:39:54 Replace packaging lettering</li><li>00:40:54 Inspect the enlarged tea label</li><li>00:41:25 Assign reference roles</li><li>00:42:37 Generate the combined campaign image</li><li>00:43:30 Layout boxes and Modern Node Design</li><li>00:44:35 Draw four placement regions</li><li>00:45:32 Match object mentions to box labels</li><li>00:46:16 Inspect all four climates and labels</li><li>00:47:12 Load ten references in order</li><li>00:48:37 Set reference resolution and caption</li><li>00:49:23 Check all ten cabinet objects</li><li>00:50:20 Create three product views</li><li>00:51:27 Images per prompt and batch results</li><li>00:52:46 Prepare a folder of text prompts</li><li>00:53:52 Queue and inspect both outputs</li><li>00:54:36 Compare prompt weighting</li><li>00:55:37 Seeded choices, schedules and BREAK</li><li>00:56:47 Negative prompts and CFG 1</li><li>00:57:18 Repeat a fixed seed</li><li>00:58:21 Randomize the composition</li><li>00:58:55 Variation seed and strength</li><li>01:00:08 Configure a compatible LoRA</li><li>01:01:07 Reuse models with extra paths YAML</li><li>01:01:50 Move the decoder and save its mapping</li><li>01:03:13 Restart and verify shared models</li><li>01:03:41 Edit Default BAT launcher arguments</li><li>01:04:27 Add — reserve-vram 2</li><li>01:05:34 Restart and verify the changed launcher</li><li>01:06:15 Cache and low-VRAM options</li><li>01:07:13 Remove or restore an argument</li><li>01:08:31 Optimized launcher and memory menu</li><li>01:09:31 Choose caching and verify launch mode</li><li>01:10:51 Update while preserving your work</li><li>01:11:33 Update ComfyUI and the node bundle</li><li>01:12:52 Test the updated setup</li><li>01:14:21 Continue with your reusable installation</li></ul><p>For anyone building a reusable ComfyUI setup. Get help through the pinned comment, Patreon, its Discord link or the comments.</p><h3>18 October 2026 Update V143</h3><ul><li>17 October 2026 Massively Detailed Windows Tutorial : <a href="https://youtu.be/yOj9PYq3XYM">https://youtu.be/yOj9PYq3XYM</a></li></ul><p>Presets Folder Content In Latest Zip File</p><p>This is a MOSS-Image update. V143 adds six new presets for MOSS-Image, the new open text-to-image model from OpenMOSS: Text To Image with the official 50-step quality recipe, Fast 25 Steps, Turbo 8 Steps, Layout Boxes, and Base model versions of Text To Image and Layout Boxes. It also adds a MOSS-Image prompt guide for ChatGPT, Claude and Gemini, a ComfyUI notebook for Google Colab, an updated AvatarForever preset with a live preview panel, fixes for the VRAM-optimized launcher and the installer, the newest ComfyUI with native MOSS-Image support, and the matching custom-node and SwarmUI extension updates.</p><p>Four native 1920×1088 images from the new Advanced Text To Image preset, with no upscaling: 50 steps, 32 to 41 seconds each on an RTX 5090. Every label and poster line came out exactly as written, and the chalkboard is spelled correctly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JIL5ZQUjZbWLd6RK.jpeg" /></figure><p>Every result in this post comes from a real run of the new presets with their default settings; only the prompt (for the layout, also the boxes), seed and output size were changed. The timings were measured on our RTX 5090 after the models were loaded.</p><p>Here is the complete Presets folder of V143: 96 ready-to-use ComfyUI presets, plus 3 API presets in the api_paid folder. The six new MOSS-Image presets are marked NEW and the updated AvatarForever preset is marked UPDATED:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MKaBHO3TCJo7FGwO.png" /></figure><p>The presets cover image, video, lip sync, audio and music models: AvatarForever, ERNIE Image, FLUX.1 Dev, Kontext and SRPO, FLUX 2, FLUX 3 Klein, Ideogram 4, Krea 2, LTX 2, LTX 2.3, LTX 2.5, MiniMax H3, MiniMax Music 3, MOSS-Image, Phantom Wan, Qwen Image, SDXL, Stable Video Infinity 2 Pro, Wan 2.1, Wan 2.2 and Z-Image.</p><h3>NEW PRESETS — MOSS-Image: Text To Image, Fast, Turbo, Layout Boxes and Base</h3><ul><li><em>“MOSS Image — Advanced Text To Image — 261017”</em></li><li><em>“MOSS Image — Fast 25 Steps — 261017”</em></li><li><em>“MOSS Image — Turbo 8 Steps — 261017”</em></li><li><em>“MOSS Image — Advanced Layout Boxes — 261017”</em></li><li><em>“MOSS Image Base — Advanced Text To Image — 261018”</em></li><li><em>“MOSS Image Base — Advanced Layout Boxes — 261018”</em></li></ul><p>Here are the new and updated presets, with what each one does:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*m-JPRV1H-LPtNX6z.png" /></figure><p>MOSS-Image is a new open image model from OpenMOSS: a 5B diffusion transformer that reads your prompt with the Qwen3.5–9B language model and uses the FLUX 2 VAE. It renders clear, accurate text inside images, accepts long and detailed prompts of up to about 4,000 tokens, and, as an experimental feature of the model, can place objects, characters and text in boxes you draw. It is released under the Apache 2.0 license, which also allows commercial use. It makes images from text; it does not edit images or use reference images.</p><p>SFT and Base: SFT is the official fine-tune of Base on curated data, and the four main presets use it; in the official comparison it groups text more clearly. Base is the raw model without post-training, with more varied images, meant for LoRA training; two Base presets are included.</p><p>All six presets share the same easy controls:</p><ul><li>One prompt box: plain language or tags, and in the Text To Image presets also a JSON layout. Long, detailed prompts work best. Put exact text for signs, labels and posters in quotes, and use real line breaks to control the lines.</li><li>Output size: choose an aspect ratio and megapixels, or type the width and height. Any aspect ratio works; 1 to 2 megapixels is recommended. Sizes snap to multiples of 16.</li><li>Batches: + Prompt adds more prompts to one run, a folder of .txt files runs a whole prompt list (JSON layouts too), and Images per prompt makes several images of each prompt.</li><li>Prompt syntax: random choices {red|blue} follow the seed in the main prompt, added prompts and folder prompts, so a fixed seed repeats the same prompt. Schedules [red:blue:0.5] and alternation [red|blue] work in plain-text prompts of the Text To Image presets. Weights like (word:1.2) are removed because MOSS-Image’s text encoder would read them as words; describe emphasis in words instead.</li><li>JSON layouts in the Text To Image presets: a complete MOSS-Image layout is sent unchanged, only without extra spaces. Pixel-based layout drafts from the official MOSS prompt enhancer or our prompt guide are converted to the model’s format automatically and scaled to your output size. In the Layout Boxes presets, the prompt box holds only the caption of the whole image.</li><li>Automatic SFT de-band: the SFT model can draw faint vertical color stripes on skies, fog and plain walls. The presets apply a community fix to SFT automatically and skip it for Base.</li><li>Seed: fixed, or randomized on every run, with an optional variation seed and strength.</li><li>Optional LoRA stack for MOSS-Image LoRAs, off by default.</li><li>Attention set to PyTorch: it matched the official MOSS-Image images in our tests, and it also prevents black images with the BF16 files when ComfyUI starts with Sage attention, as our launchers do (RTX 3090 test). Keep it on PyTorch in your own BF16 workflows.</li><li>The model, text encoder, VAE and latent are already connected. To use the BF16 files, switch the model loader. Use Base only in the 50-step presets: the Fast and Turbo recipes were made for SFT.</li></ul><p>Here is the Advanced Text To Image preset after a real run of a magazine cover prompt:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cZ3pFwQJir3UUvlF.png" /></figure><p>All five text lines came out exactly as written, from the huge WANDER title to the small issue line. The 1024×1536 image took 21.8 seconds with the official 50-step recipe.</p><p>Fast 25 Steps, Turbo 8 Steps and Base: here is the same prompt and seed through all four text-to-image presets at 1920×1088:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tnkukRO_oCCx6TkP.jpeg" /></figure><ul><li>Advanced Text To Image (SFT, 50 steps, Euler / normal, CFG 3, shift 5, the official recipe): the most faithful results, especially for text and designs. 34.3 seconds here, 12.9 seconds at 1024×1024.</li><li>Fast 25 Steps (ER-SDE / simple, CFG 3, Temporal Score Rescaling 0.95, layer skip 22 and 24 from 10% to 70% of the steps): 16.6 seconds here, about half the time of the 50-step preset. Large text holds up well; very small text can occasionally lose a letter.</li><li>Turbo 8 Steps (the community 8-step LoRA by lodestones at 0.7, Euler / simple, CFG 1): 3.4 seconds here, about a tenth of the 50-step time. The 8-step LoRA is a work in progress, so its images differ more from the 50-step ones and can show extra signs.</li><li>Base Advanced Text To Image (Base model, the official 50-step recipe): more varied images; 32.5 seconds here. For text-heavy designs, start with SFT: in the official Base versus SFT example, SFT groups text more clearly.</li></ul><p>Layout Boxes: draw a box for each object, character or line of text and describe it. The main prompt becomes the caption of the whole image:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*x_STb17nIefO8qsz.png" /></figure><ul><li>Each box is an object or exact text, with its own description. To turn object boxes into characters, type their numbers in character_boxes, here 8 for the barista.</li><li>The canvas follows the output size, and the preset builds MOSS-Image’s layout JSON from your boxes.</li><li>The box editor needs Settings → Modern Node Design (Nodes 2.0). Turn it on once.</li><li>Boxes guide placement. They are not masks, so positions can vary a little.</li></ul><p>Here is the result of that layout in the same preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*34UUY0uDbyxxTeMR.png" /></figure><p>Every box landed where it was drawn: the headline, three drinks with their exact names and prices, and the waving barista as a character. The 1536×1024 image took 22.6 seconds.</p><p>MOSS-Image, the 8-step LoRA and the de-band fix are released under the Apache 2.0 license.</p><h3>UPDATED PRESET — AvatarForever: Live Preview Panel and Tested Mouth-Movement Settings</h3><p><em>“AvatarForever Unified Image or Text + Audio To Long Video — 260920”</em></p><ul><li>New Live Preview &amp; Final Video panel to the right of the generation controls. It shows a frame preview after each sampling step, starting with the first, then plays the finished MP4.</li><li>New tested preset settings: a 512×768 canvas and 16 latent frames per chunk, instead of 832×1248 and 4. In our singing demo, 4-frame chunks kept the lips closed, while 16-frame chunks gave sustained mouth movement through the whole 18-second song. At 1024×1024 the mouth stayed mostly closed even with 16-frame chunks, so check a short sample when you change the image, audio, resolution or chunk size. Larger chunks need more memory; 4-frame chunks still work for speech and use less.</li><li>Sharper previews: the live preview uses the small taeltx2_3 preview decoder when it is installed. It is now in the LTX 2.3 Avatar Forever Core Bundle of our SwarmUI Model Downloader V183. Without it, the preview uses a coarser latent view.</li><li>Your own AvatarForever workflows keep working and now show live previews on the sampler. Set live_preview to off to turn them off.</li></ul><h3>NEW PROMPT GUIDE — MOSS-Image Feed for ChatGPT, Claude and Gemini</h3><p>A new MOSS_Image_Enchance_Prompt_Feed_For_LLMs.txt is included. Give it to ChatGPT, Claude or Gemini together with your request, and it writes prompts that MOSS-Image understands:</p><ul><li>Detailed captions for photos, illustrations and designs, with the exact visible text in quotes and where and how it appears.</li><li>Layouts as pixel-based JSON drafts for the Text To Image presets, which convert them to MOSS-Image’s own format automatically, or a box-by-box plan to draw in the Layout Boxes editor.</li><li>Tag-style prompts for anime, in the recommended tag order.</li><li>The right settings: output size, CFG for short prompts, the negative prompt and when to use SFT or Base.</li></ul><h3>NEW — ComfyUI Colab Master Notebook for Google Colab</h3><p><em>“ComfyUI_Colab_Master_Notebook.ipynb”</em></p><p>A new Google Colab notebook runs ComfyUI with ComfyUI-Manager on a Colab GPU and opens it in your browser through a Cloudflare link. It sets up Qwen-Image 2.1 for you:</p><ul><li>Install: ComfyUI (the newest stable release, or master if you choose it) with ComfyUI-Manager and cloudflared, which makes the link without an account. Colab’s own PyTorch is kept.</li><li>Qwen-Image 2.1 in the right size: the official ComfyUI files are picked from your GPU memory and system RAM: INT8 with the W4A8 text encoder (14.2 GB) when either is below 20 GiB, as on a T4; INT8 with the INT8 text encoder (17.3 GB), the official template’s files, when both have at least 20 GiB; or full BF16 (32.4 GB) when you choose it. Broken downloads resume when you run the cell again.</li><li>Any other model: paste up to four direct download links, such as LoRAs or upscalers, and choose the folder or let the notebook read it from the link.</li><li>Start: you get a <a href="https://trycloudflare.com/">trycloudflare.com</a> link, and ComfyUI opens with the Qwen-Image 2.1 text-to-image workflow already loaded, with the official settings (Euler, simple, 25 steps, CFG 1, 1024×1024).</li><li>Keep your work: one cell packs your outputs and saved workflows into a ZIP to download. The notebook also explains how to upload images and videos, and what is kept when the runtime ends.</li></ul><p>Open Google Colab, choose Upload notebook, select the file, pick a GPU runtime (T4 or better) and run steps 1 to 4 in order. T4 GPUs can be allocated on free accounts, but availability is not guaranteed, and Colab says free runtimes used mainly through an external web UI may be terminated. The Cloudflare link is public, so don’t share it. The Qwen-Image 2.1 weights use the Qwen Research License, which is non-commercial.</p><h3>LAUNCHER AND INSTALLER FIXES</h3><ul><li>Windows_Run_VRAM_Optimized.bat: option 3 (normal VRAM), and VRAM choice 3 in Custom mode, passed — normalvram, which ComfyUI removed in version 0.22.0, so ComfyUI stopped at once with “unrecognized arguments: — normalvram”. Both now start ComfyUI with its default memory management and no extra flag.</li><li>Windows_Run_VRAM_Optimized.bat also describes the memory modes as current ComfyUI uses them. LOWVRAM’s — lowvram has no effect while dynamic VRAM is on, the default on NVIDIA GPUs, and otherwise only runs the text encoders on the CPU; NOVRAM turns dynamic VRAM off. Pressing Enter in the cache menu now keeps ComfyUI’s own default, RAM-pressure caching, instead of the older classic caching; choose 1 for classic caching.</li><li>Windows_Install_Or_Update_ComfyUI.bat: the pip update and the uv install now ignore your personal pip settings (the pip.ini in your user profile and PIP_ environment variables), so a custom package index set there for another project can no longer stop the installer at this step. Its Windows requirements tutorial link now points to the current video: <a href="https://youtu.be/ptuLL0C72Ms">https://youtu.be/ptuLL0C72Ms</a></li></ul><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V143.zip from this post.</li><li>Extract it over your existing installer folder and overwrite the older files.</li><li>Close ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat. It updates ComfyUI to the newest version, with MOSS-Image support. The MOSS-Image presets do not run on older ComfyUI.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install/update bundle 100. It updates SECoursesAudioTools with the new MOSS-Image nodes and the AvatarForever Live Preview &amp; Final Video node that the updated AvatarForever preset uses, and keeps FoleyExtension, SwarmUI ExtraNodes and ComfyUI-Easy-Use, which the MOSS-Image presets also use, up to date.</li><li>Restart ComfyUI and refresh the browser tab.</li><li>For the Layout Boxes editor, turn on Settings → Modern Node Design (Nodes 2.0).</li></ol><p>Use our latest SwarmUI Model Downloader (SwarmUI_Model_Downloader_v183.zip or newer) and download the MOSS-Image Core Bundle, the first card of the SwarmUI Bundles tab. It has the six files the six presets use, about 23.2 GB in total (the downloader counts in GiB, so it shows 21.64 GB):</p><ul><li>moss_image_sft_int8_convrot.safetensors (5.41 GB), the SFT model in INT8 ConvRot.</li><li>moss_image_base_int8_convrot.safetensors (5.41 GB), the Base model in INT8 ConvRot.</li><li>qwen3.5_9b_int8_convrot.safetensors (9.75 GB), the Qwen3.5–9B text encoder in INT8 ConvRot.</li><li>Flux/flux2-vae.safetensors (0.34 GB), the FLUX 2 VAE, the same file the FLUX 2 and FLUX 3 presets use.</li><li>MOSS_Image_SFT_DeBand_L37–41.safetensors (1.13 GB), the SFT de-band fix.</li><li>MOSS_Image_SFT_8_Steps_DMD_LoRA.safetensors (1.19 GB), the 8-step LoRA of the Turbo preset.</li></ul><p>The INT8 ConvRot model files are about half the size of BF16, and in our RTX 3090 test they were about twice as fast (48 versus 99 seconds at 50 steps). The BF16 models, a smaller W4A8 text encoder (select it in the text encoder loader), the small-decoder FLUX 2 VAE and the extra de-band fix for layers 32–36 are optional downloads under Image Generation Models → MOSS-Image Models. The de-band author recommends the extra fix for images 1536 pixels wide or larger and for big, smooth skies; select it in the de-band node.</p><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>17 October 2026 Update V142</h3><p>This is a FLUX 3 Klein update. V141 adds three new presets for Black Forest Labs’ new FLUX 3 Image Klein model: Text To Image, Image Edit with up to 10 reference images, and Layout Boxes. It also adds a FLUX 3 Klein prompt guide for ChatGPT, Claude and Gemini, our new HQ INT8 ConvRot FLUX 3 Klein model, ComfyUI 0.41.0 with native FLUX 3 support, and the matching custom-node and SwarmUI extension updates.</p><p>A native 3840×2160 image from the new Text To Image preset, with no upscaling: 16 steps in 23 seconds on an RTX 5090.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nZWPNvqLVYzl0eVI.jpeg" /></figure><p>Every result in this post comes from a real run of the new presets with their default 16 steps, Euler / simple, CFG 1. The timings were measured on our RTX 5090 after the models were loaded.</p><h3>NEW PRESETS — FLUX 3 Klein: Text To Image, Image Edit and Layout Boxes</h3><p><em>“FLUX 3 Klein — Advanced Text To Image — 261015”</em></p><p><em>“FLUX 3 Klein — Advanced Image Edit — 261015”</em></p><p><em>“FLUX 3 Klein — Advanced Layout Boxes — 261015”</em></p><p>FLUX 3 Image Klein is a 5.6B image model with the Qwen3-VL-4B text encoder. One model does text to image, image editing, up to 10 reference images, layout boxes and text rendering. Our presets run the fast distilled model with its recommended 16 steps and no guidance (CFG 1), using Euler / simple.</p><p>All three presets share the same easy controls:</p><ul><li>One prompt box with an image gallery: write the whole scene. Add up to 10 images and refer to them as @image1 … @image10 (@img1, @picture1 and @pic1 also work). With no image, it is text to image.</li><li>Output size: choose an aspect ratio and megapixels, or type the width and height. Sizes snap to multiples of 16.</li><li>Batches: + Prompt adds more prompts to one run, a folder of .txt files runs a whole prompt list, and Images per prompt makes several images of each prompt.</li><li>Advanced prompt syntax: weights (ceramic:1.1), random choices {red|blue}, schedules [red:blue:0.5], alternation [red|blue] and &lt;break&gt;. Random choices follow the seed, so a fixed seed repeats the same prompt.</li><li>Seed: fixed, or randomized on every run, with an optional variation seed and strength.</li><li>Reference size: one Reference megapixels setting resizes every reference image and keeps its shape (1 MP by default, up to 4 MP). It is separate from the output size.</li><li>Optional LoRA stack for FLUX 3 LoRAs, off by default.</li><li>The model, text encoder, VAE and latent are already connected. There is nothing else to set up.</li></ul><p>Here is the Text To Image preset after a real run of a poster prompt with English and Turkish text:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xuiD3RycIRASgFOH.png" /></figure><p>All four text phrases came out exactly as written, including the Turkish characters. The 1024×1536 image took 3.2 seconds. The first run takes longer because it loads the models.</p><p>Image Edit: the same workflow with your image in the gallery. Here it turns a celadon museum vase cobalt blue and keeps its shape, handles, framing and shadow:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ab7tjZ-bp8HEk0Tk.png" /></figure><ul><li>Say what changes and what stays, and refer to your photo as @image1.</li><li>Match the output aspect ratio to your photo to keep its framing. This 1024×1536 edit took 5.2 seconds.</li><li>The preset opens with a sample card. Add your own image with + Add reference images and remove any card marked “file missing”.</li></ul><p>Up to 10 reference images in one prompt: here ten separate museum photos became one display wall, each object in its requested niche:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RKQCOcyX5vE0mEIs.png" /></figure><p>All ten objects appear in the requested order and closely match their photos. The 1680×944 image took 41 seconds with all ten references at 1 MP each. The reference photos are public domain (CC0) images from The Cleveland Museum of Art.</p><p>Layout Boxes: draw a box for each object or line of text, describe it, and refer to it in the caption as &lt;obj_1&gt;, &lt;obj_2&gt; and so on:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-bnbduqu63q9ko47.png" /></figure><ul><li>Each box is an object or exact text, with its own description. The preset builds FLUX 3’s layout table from your boxes, and the boxes follow the output size.</li><li>The box editor needs Settings → Modern Node Design (Nodes 2.0). Turn it on once.</li><li>Boxes guide placement. They are not masks, so positions can vary a little.</li></ul><p>Here is the result of that layout in the same preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*txxFZHWrnP7bVxgG.png" /></figure><p>All four terrariums and captions landed where their boxes are, and the four captions are spelled exactly. The 1024×1024 image took 2.5 seconds.</p><p>FLUX 3 Klein is released under the FLUX Kommunity License: free for non-commercial use, and people and companies with a yearly revenue under US$5 million (affiliates included) may use the images commercially under its conditions. Read the license for the full terms.</p><h3>NEW MODEL — Our FLUX 3 Klein HQ INT8 ConvRot</h3><p><em>“FLUX-3-Image-Klein-INT8-ConvRot-HQ.safetensors”</em></p><p>The presets use our own INT8 ConvRot conversion of the distilled FLUX 3 Klein model. On the updated ComfyUI it loads with the normal Load Diffusion Model node, so no special loader is needed.</p><ul><li>6.07 GB instead of 11.1 GB for the original BF16 model. While denoising at up to 1024×1024, the model alone peaked at about 6 GiB of VRAM instead of about 10.6 GiB for BF16 (RTX 3090 test).</li><li>About 2× faster per step than BF16 (1.9–2.3× in our RTX 3090 tests, model only).</li><li>Closer to BF16 than the official INT8 file: in our test of 40 sampling states, its mean output error against BF16 was 2.14%, versus 3.05% for the official INT8 ConvRot file with its default settings and 2.27% with matched attention. One held-out editing case favored the official file with matched attention. In exchange, our file is 0.25 GB larger and was 6–22% slower per step in the same RTX 3090 test. These are numerical measurements made with the BF16 text encoder, not image-quality percentages.</li><li>The first text block and the last joint block, the most sensitive in our tests, stay in BF16, as do the small conditioning, normalization and output layers. The other 164 linear layers use native INT8 ConvRot.</li></ul><h3>NEW PROMPT GUIDE — FLUX 3 Klein Feed for ChatGPT, Claude and Gemini</h3><p>A new FLUX_3_Klein_Enchance_Prompt_Feed_For_LLMs.txt is included. Give it to ChatGPT, Claude or Gemini together with your request and images, and it writes prompts that FLUX 3 Klein understands:</p><ul><li>Complete scene captions for text to image, with the most important details first.</li><li>Edit prompts that say what changes and what must stay, and multi-reference prompts that give every image a clear role.</li><li>Exact text for posters and designs, in the original language and spelling.</li><li>Layout prompts in the FLUX 3 box format, with the &lt;obj_1&gt; labels these presets use.</li><li>The right settings and the ComfyUI / SwarmUI details, such as @image1 meaning FLUX 3’s first reference image.</li></ul><h3>AvatarForever Preset Now Has Live Preview Feature</h3><ul><li>2:3 and 3:2 aspect ratios works best — 1:1 not that good</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hkkXGcXtMZa4bHYU.png" /></figure><h3>HOW TO UPDATE</h3><ol><li>Download the latest ComfyUI_V141.zip from this post.</li><li>Extract it over your existing installer folder and overwrite the older files.</li><li>Close ComfyUI, then run Windows_Install_Or_Update_ComfyUI.bat. It updates ComfyUI to the newest version, with FLUX 3 support. The FLUX 3 presets do not run on older ComfyUI.</li><li>Run Windows_Custom_Nodes_Bundles_Installer.bat and install/update bundle 100. It updates SECoursesAudioTools with the new FLUX 3 nodes and keeps FoleyExtension, SwarmUI ExtraNodes and ComfyUI-Easy-Use, which the presets also use, up to date.</li><li>Restart ComfyUI and refresh the browser tab.</li><li>For the Layout Boxes editor, turn on Settings → Modern Node Design (Nodes 2.0).</li></ol><p>Use our latest SwarmUI Model Downloader (SwarmUI_Model_Downloader_v181.zip or newer) and download the FLUX 3 Klein Core Bundle, the first card of the SwarmUI Bundles tab. It has the three files all three presets use, about 11.7 GB in total (the downloader counts in GiB, so it shows 10.85 GB):</p><ul><li>FLUX-3-Image-Klein-INT8-ConvRot-HQ.safetensors (6.07 GB), our HQ INT8 ConvRot model.</li><li>qwen3vl_4b.safetensors (5.24 GB), the FP8 Qwen3-VL-4B text encoder.</li><li>Flux/flux2-vae.safetensors (0.34 GB), the FLUX 2 VAE, which FLUX 3 uses too.</li></ul><p>The text encoder is the same file Krea 2 uses and the VAE is the FLUX 2 VAE, so if you already have those, only the 6.07 GB model is new.</p><p>SwarmUI Installer and Unified Massive Models Downloader: <a href="https://www.patreon.com/posts/114517862">https://www.patreon.com/posts/114517862</a></p><h3>10 October 2026 Update V140</h3><p>ComfyUI V140 is ready. Here is everything we changed since V138: 4 new MiniMax H3 presets, optional MiniMax H3 RefMods (reusable reference files) in our MiniMax H3 presets, a native sparse attention engine in our MiniMax H3 Speed Optimizer, fixes that keep our nodes working on the newest ComfyUI (tested on ComfyUI 0.39.0), and fixes for every preset that failed when we ran all presets with real generations.</p><h3>New MiniMax H3 Presets</h3><p>MiniMax H3 Lip Synch — Turbo 4 Steps (672x1216 and 832x1248)</p><p>Here is the 672x1216 preset after a real run with its demo audio and portrait:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ilmoXm0joptr3jRJ.png" /></figure><ul><li>Fast talking videos from your audio: upload an audio file, optionally add an image, describe the scene and run. The video length follows your audio. Without an image it works as audio + text. With a video in the reference gallery, it makes a new performance from that video.</li><li>Turbo 4 steps by default with the LightX2V FL2VA Turbo 4-step v1.2 LoRA, at 24 fps. A 15-second video took about 165 seconds at 672x1216 on an RTX 5090. Set “4. QUALITY / SPEED / VRAM” to FULL QUALITY for the 50-step path.</li><li>Two portrait resolutions: 672x1216, or 832x1248 for more detail. These are 4-step portrait versions of our existing MiniMax H3 Lip Synch preset (1024x1024, 8 steps).</li></ul><p>The 832x1248 preset saves the original video first, then runs the new CodeFormer mouth pass:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pkJYc6DD85oZfTvq.png" /></figure><ul><li>The 832x1248 preset adds a CodeFormer mouth pass (on by default). It restores only the mouth area of the finished video and saves it as a second MP4, next to the original video.</li><li>Also included: our optional face refinement (off by default) and the new optional RefMods.</li><li>Tip: for your own audio, put its exact words in the &lt;d&gt;[English] … &lt;/d&gt; part of the scene prompt, or clear the scene prompt to use the automatic one. 5–15 seconds per generation is recommended; longer audio still works in one pass.</li></ul><h3>MiniMax H3 Long Lip Synch — Native Latent Continuation — Full Audio</h3><p>Here is the Long Lip Synch preset with its default settings:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tUUZrZpx0VuY6nv_.png" /></figure><ul><li>One full soundtrack in, one continuous video out: give it a complete audio file plus a portrait image or gallery references. It makes one continuous 24 fps lip-synced video as long as your audio. There is no duration setting and no cutting.</li><li>How it stays continuous: it generates overlapping windows of about 10 seconds (243 frames) and carries the last 1.6 seconds (39 frames) of each window forward as clean latents, so nothing is re-encoded between windows. The audio is encoded once, and your original soundtrack is put back into the final MP4.</li><li>Works with references and RefMods in every window, unlike the TaoMate streaming preset. The first-frame image anchors only the start.</li><li>Defaults: 832x1248 and Turbo 4 steps (FL2VA 4-step v1.2, or Ref2VA 4-step v0.1 when you use references). Use 512x768 for faster drafts. SOL sparse attention and FirstBlockCache are separate switches, both off by default.</li><li>Outputs: the final MP4 with your original audio, plus a JSON report with the timing of every window. On an RTX 5090, 30 seconds of 512x768 video took 105 seconds with normal attention.</li><li>Experimental: very long videos can change framing over time. In our 15-clip test, 13 clips passed; the 45 and 60-second clips at 832x1248 drifted into close-ups.</li></ul><h3>MiniMax H3 RefMod Builder — Images, Video, Audio</h3><p>Add your source media, name your RefMod and run. Only the MiniMax H3 VAEs are used:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7V0CS8fBEI6IwiMu.png" /></figure><ul><li>Build your own RefMod in one run: add photos of one subject, a video (its soundtrack is included), audio, or a mix. Name it and run. Only the MiniMax H3 VAEs run: no diffusion model, no text encoder and no training.</li><li>Options: source resolution (512 by default), “one character” (stacks the photos into one reference) or “separate references” (each photo gets its own picture label), and optional compression (32, 16 or 8) for fewer tokens at the cost of detail.</li><li>Ready to use: files are saved to ComfyUI/models/refmods and existing files are never overwritten. Pick the file in any MiniMax H3 preset’s Optional RefMods node and write @refmod1 in your prompt.</li></ul><h3>MiniMax H3 RefMods: Reusable References</h3><p>A RefMod is a small .safetensors file that stores already-encoded MiniMax H3 references: images, video and/or audio of a character, scene or voice. Build it once, then reuse or share it in any MiniMax H3 preset without encoding the source media again. It is not a LoRA and needs no training.</p><p>Here is the new Optional RefMods node with an imported RefMod, used as @refmod1 in the prompt:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LQcXW1bN18IQqqBp.png" /></figure><ul><li>New “MiniMax H3 Optional RefMods” node in all 9 existing MiniMax H3 generation presets (Text To Video, Image or Video To Video, References To Video, Audio Only, Text With References To Image and Lip Synch) and in the new presets. It is empty by default, so your results do not change unless you add a RefMod.</li><li>Easy picker: “+ Add RefMod”, “Import files” and “Refresh” buttons. Each row has an on/off box, the file (with its contents, for example “image + audio”), a strength from 0 to 1, a visual/audio filter and up/down ordering. Imported files go to ComfyUI/models/refmods and are checked before use.</li><li>Use them in your prompt with @refmod1, @refmod2 and so on. They are numbered after your gallery images, videos and audio.</li><li>Strength sets how much reference detail is kept. It is not a LoRA weight.</li><li>RefMods alone are enough: the reference mode and auto mode nodes count RefMods as references, so a RefMod without any gallery image switches generation to the references (Ref2VA) path.</li><li>Compatible files: RefMods made with ComfyUI-MiniMaxH3Mod and the Fantastic MiniMaxH3 PromptBuilder load too, without installing those node packs. Their MIT-licensed format code is credited in our notices.</li><li>TaoMate streaming does not use RefMods; its preset note now points you to the new Long Lip Synch preset for RefMods with long audio.</li></ul><h3>MiniMax H3 Speed Optimizer: Native Sparse Attention</h3><p>Here is the MiniMax H3 Speed Optimizer node, shown in the Text With References To Image preset, where it is off by default:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_adITtifhA2f573r.png" /></figure><ul><li>SOL sparse attention now runs on ComfyUI’s own MiniMax H3 sparse attention (our ComfyUI-TeaCache fork). The old bundled attention kernels are gone, so there is nothing extra to compile.</li><li>Faster than our old SOL: the old version repeated work and was slower than normal attention at 832x1248 (207 s against 116 s). With the new default protected SOL settings, a 10-second 832x1248 Turbo 4-step video took 99 seconds instead of 111 seconds with normal attention (1.12x, RTX 5090, FirstBlockCache off for this test). At 512x768 the gain was 1.06x. Sparse attention and caching are approximations, so check the quality of your own results.</li><li>Auto mode measures before it switches: it uses sparse attention only when it is at least 5% faster. The Text To Video and Image or Video To Video presets have the optimizer on in auto mode, so they now use the native path automatically when it pays off.</li><li>FirstBlockCache works together with other speed-ups: it now keeps patches that are already on the model, such as the sparse attention added by the lip-sync Turbo node, instead of replacing them.</li><li>Old presets load unchanged. The main switch is now called “H3 optimizations” instead of “4x speedup”, and “native” is the new default backend.</li></ul><h3>Other MiniMax H3 Improvements</h3><ul><li>Correct colours in streamed MiniMax H3 videos: TaoMate streaming and long lip-sync videos are now tagged with the colour format they are encoded in (BT.601 limited range), so players read the colours correctly.</li><li>No torchaudio needed for MiniMax H3 audio: init audio, streaming and long lip sync now resample any soundtrack with ComfyUI’s own resampler, with bit-identical results (details below).</li></ul><h3>Works With the Latest ComfyUI</h3><p>ComfyUI removed torchaudio and changed its attention and Linear layer code. We updated our nodes so everything keeps working on the newest ComfyUI:</p><ul><li>No more torchaudio needed: our MiniMax H3 streaming and long lip-sync nodes (SECoursesAudioTools), the init audio nodes (FoleyExtension) and the audio models in our ComfyUI-WanVideoWrapper fork (HuMo, Ovi, FantasyTalking, MultiTalk and LongCat) now resample audio with ComfyUI’s own resampler. Before, on installs without torchaudio, they failed on any soundtrack that was not already at the model’s sample rate, and HuMo and Ovi did not load at all. The output is bit-identical: 88 test comparisons had zero difference, and a real MiniMax H3 Turbo 4 video was pixel-identical on all 120 frames.</li></ul><p>Here are the new torchaudio-free Mel-Band RoFormer nodes inside the LTX2.3 audio lip-sync preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*P4DGcbPxBDkyH-h-.png" /></figure><ul><li>New torchaudio-free Mel-Band RoFormer vocal separation nodes (SECoursesAudioTools: “SE Mel-Band RoFormer Model Loader” and “SE Mel-Band RoFormer Sampler”). They load the same checkpoint and give bit-identical vocals, so the LTX2 and LTX2.3 audio lip-sync presets keep their vocal separation even where the original Mel-Band RoFormer node pack cannot load.</li><li>GGUF models work again (our ComfyUI-GGUF fork): on the latest ComfyUI every quantized GGUF model failed with “unexpected keyword argument ‘input_act’”. Fixed and verified bit-exact on 57 test cases.</li><li>FLUX 2 Mistral GGUF text encoder loads again (ComfyUI-GGUF): the GGUF Tekken tokenizer now gives ComfyUI the special token count it needs. Token IDs match the official tokenizer.</li><li>Much faster FLUX 2 GGUF on limited VRAM (ComfyUI-GGUF): GGUF text encoders are now fully unloaded when VRAM is needed, instead of leaving 6.6 GB in VRAM. FLUX 2 Low RAM sampling went from 200–309 s/it to 1.75 s/it on a 32 GB GPU.</li><li>LTX 2.3 Foley presets fixed: ComfyUI-LTXVideo is now installed from our fork, which fixes the STG error “‘AttentionTensorContainer’ object has no attribute ‘shape’”. Existing installs switch to the fork automatically on their next update.</li><li>ControlNet Aux keeps the MIT albumentations package (our comfyui_controlnet_aux fork): upstream switched to AlbumentationsX, which is AGPL-3.0, sends usage telemetry by default and conflicts with insightface (used by ReActor).</li></ul><h3>Preset Fixes and Updates</h3><p>We ran every preset with real generations on the latest ComfyUI and fixed every one that failed. The FLUX 2 Low RAM presets now free VRAM between prompt encoding and sampling:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*d9jx7Ow1x0tYd0dO.png" /></figure><ul><li>FLUX 2–8 Steps Fast HQ / Realism — Low RAM: two new “Free VRAM before sampling” nodes unload the models after the prompts are encoded. Speed went from 200–309 s/it to 1.75 s/it, 35–75 seconds per image on a 32 GB GPU. The Realism Low RAM preset also had the HQ preset’s guidance of 7; it is now 1, like the normal Realism preset.</li><li>Krea 2 Raw Base, Krea 2 Turbo and Krea 2 Turbo Image Edit: the saved sampler values were shifted by one position (for example steps became 1 and CFG became “euler”). Fixed.</li><li>LTX2 Image To Video and LTX2 Text To Video 8 Steps: the seed was empty, so generation failed. Fixed.</li><li>LTX2.3 Extend Any Video With Audio: the input video is now resized to a multiple of 64, so the source and the half-resolution second stage always match (1280x720 is unchanged).</li><li>LTX2.3 Foley Video To Audio: the video loader read the frame limit (169) as “skip first frames”. Fixed.</li><li>LTX2.5 Text To Audio: the text encoder and prompt enhancer were connected to the wrong inputs, and the enhance prompt option held a model file name. Fixed.</li><li>LTX2.5 IC LoRA Pixel Spatial Upscaler 2x (FAST Distilled and Slower HQ): the VAE paths now point to the LTX-2 folder, where our model downloader saves them.</li><li>LTX2 and LTX2.3 Audio Lip Synch Image To Video (3 presets): now use our new torchaudio-free Mel-Band RoFormer nodes for vocal separation.</li><li>AvatarForever Unified Image or Text + Audio To Long Video: now renders an exact 832x1248 portrait by default, instead of a 768x512 pixel budget that followed the input image’s shape.</li></ul><h3>Installer Improvements</h3><ul><li>Smarter node updates (Windows, RunPod and Massed Compute): the updater now asks GitHub for each node’s default branch and fetches only that branch. When we move a node to a different fork, as with ComfyUI-LTXVideo in this release, existing installs switch over cleanly.</li><li>Reviewed fork updates: our forks of rgthree-comfy, ComfyUI-Easy-Use, comfyui_controlnet_aux and RES4LYF are synced with their latest upstream code. We read every new upstream commit for safety first, then re-ran our presets with identical prompts: 85 of 85 passed.</li><li>comfy-kitchen minimum version raised to 0.2.37, which the new MiniMax H3 sparse attention and long lip sync need.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bYdWkOftEPbCEv5vr6BCQQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cc510e8a14e3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Zero to AI-Ready PC in One Click: Windows & Ubuntu AI Requirements Installer Full Tutorial]]></title>
            <link>https://medium.com/@furkangozukara/zero-to-ai-ready-pc-in-one-click-windows-ubuntu-ai-requirements-installer-full-tutorial-4cf820162dd5?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/4cf820162dd5</guid>
            <category><![CDATA[generative-ai-course]]></category>
            <category><![CDATA[image-generation]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[comfy-ui]]></category>
            <category><![CDATA[tutorial]]></category>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Sat, 17 Oct 2026 23:44:31 GMT</pubDate>
            <atom:updated>2026-10-17T23:44:31.136Z</atom:updated>
            <content:encoded><![CDATA[<p>Full tutorial link &gt; <a href="https://www.youtube.com/watch?v=ptuLL0C72Ms">https://www.youtube.com/watch?v=ptuLL0C72Ms</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FptuLL0C72Ms%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DptuLL0C72Ms&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FptuLL0C72Ms%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/57d772973020bd371029f5f6aab5006a/href">https://medium.com/media/57d772973020bd371029f5f6aab5006a/href</a></iframe><p>Full tutorial link &gt; <a href="https://www.youtube.com/watch?v=ptuLL0C72Ms">https://www.youtube.com/watch?v=ptuLL0C72Ms</a></p><p>The 1-Click Windows and Ubuntu AI Requirements Installers set up your PC for AI apps such as ComfyUI, SwarmUI, music, text-to-speech and text-to-3D tools: one click on Windows, one command on Ubuntu. They install the latest NVIDIA driver, CUDA 12.9 + 13.2, cuDNN 9.20, Python 3.10–3.14, the C++ build tools, Git, Node.js, .NET SDK and FFmpeg, set PATH in the right order, test everything for real (78 Windows and 66 Ubuntu tests passed on our test PCs) and show every result in a detailed HTML report.</p><h3>Links:</h3><ul><li>Download the installers (latest zip) and the full guide: [ <a href="https://www.patreon.com/SECourses/posts/1-click-windows-172290854">https://www.patreon.com/SECourses/posts/1-click-windows-172290854</a> ]</li><li>Previous manual method (older Windows requirements tutorial): [ <a href="https://youtu.be/DrhUHnYfwC0">https://youtu.be/DrhUHnYfwC0</a> ]</li><li>SECourses Discord for help (send your logs.txt): [ <a href="https://discord.com/invite/software-engineering-courses-secourses-772774097734074388">https://discord.com/invite/software-engineering-courses-secourses-772774097734074388</a> ]</li><li>You will learn how to extract the zip in your Downloads folder, run the Windows installer as administrator, follow its 20 steps, read the HTML report and logs.txt, run the Ubuntu installer with one sudo command, and rerun either installer to repair or update your setup.</li><li>Main topics include 1-click AI PC setup, CUDA and cuDNN, Python 3.10–3.14, Visual Studio C++ Build Tools, PATH order, torch.compile and Triton readiness, FFmpeg for torchcodec, Ubuntu 22.04/24.04/26.04 and real install tests.</li></ul><p>Use the chapters to jump to the download, the Windows steps, the HTML report, the Ubuntu installer and updates.</p><h3>Chapters:</h3><ul><li>00:00:00 One click sets up your PC for AI: all tests passed</li><li>00:00:17 Windows 10, 11 and Ubuntu 22, 24, 26: ready for any AI app</li><li>00:01:03 Download the latest zip and extract it in Downloads</li><li>00:01:19 Right-click the .bat and choose Run as administrator</li><li>00:01:38 The 20 steps begin: long paths, Store Python removed</li><li>00:01:50 CUDA 12.9 + 13.2 toolkits, then cuDNN for PyTorch</li><li>00:02:14 Visual Studio 2022 Build Tools and the VC++ runtime</li><li>00:02:43 Python 3.10 to 3.14, with 3.12 as the default</li><li>00:03:12 Git, Node.js 24, .NET SDK and FFmpeg</li><li>00:03:48 PATH order, Git long paths and the page file</li><li>00:04:09 Triton cache, NVIDIA driver and the final tests</li><li>00:04:32 The HTML report: 78 tests, every step, PATH and backups</li><li>00:05:13 Transparent source files, logs.txt, then restart Windows</li><li>00:06:05 Ubuntu 22.04, 24.04, 26.04: one sudo command</li><li>00:07:00 Rerun any time to update; GCC, Clang, CMake, Ninja</li><li>00:07:46 Ubuntu report: OK messages, warnings and logs.txt</li><li>00:08:54 One-time setup, ready for our apps and other AI apps</li><li>00:09:30 Latest installer version, the guide post and support</li></ul><p>What it installs on Windows: the latest compatible NVIDIA WHQL driver for GeForce RTX 20-series and newer (checked on every run, never downgraded); Python 3.10 to 3.14 for all users, 3.12 as default; CUDA 12.9 and 13.2 side by side (13.2 default) and cuDNN 9.20, the only cuDNN registered, so PyTorch never mixes two versions; Visual Studio 2022 Build Tools (MSVC, Windows SDK) that keep compiling after Visual Studio updates, the latest Visual C++ runtime and CUDA integration; Git, Node.js 24, .NET 10 SDK for SwarmUI, FFmpeg 9 shared and static; PATH in the right order with your own entries kept, long paths, a 100 GB page file. Ready for torch.compile, Triton and flash-attention, sage-attention, xformers and insightface builds.</p><p>On Ubuntu the C++ tools are GCC, Clang, CMake and Ninja, plus Git LFS, uv, free-threaded Python 3.13t and 3.14t and a swap file. With an NVIDIA GPU it installs the newest driver Ubuntu offers and never downgrades it. Ubuntu’s own python3 stays as it is.</p><p>Windows: put the zip in Downloads (not directly on C:), extract it, right-click Windows_AI_Requirements_Installer.bat and choose Run as administrator. Nothing has to be installed first. Ubuntu: open a terminal in the folder and run sudo bash ./Ubuntu_AI_Requirements_Installer.sh (24.04 recommended).</p><p>Afterwards the HTML report opens and logs.txt is saved next to the installer. Restart your PC and open a new terminal for the new PATH. You need this setup only once, not for every app. When the post shows a newer version, run the new one: finished parts are only checked, failed parts are repaired and components are updated, with backups first ( — restore-env puts them back).</p><p>Tested for real: the tests compile and start C++ and CUDA programs, build Python and PyTorch-style CUDA extensions, compile GPU code for RTX 50 cards, encode a video and build a .NET app, in the environment a new terminal gets; the optional PyTorch GPU test runs with — torch-test. Supported: Windows 10 (21H2+) and 11; Ubuntu 22.04, 24.04 and 26.04 (desktop, server, cloud GPU, WSL2, Linux Mint, Zorin OS), 64-bit x86. Without an NVIDIA GPU everything still installs and compiling works; CUDA apps need an NVIDIA GPU.</p><p>This video is for anyone who wants to run AI apps locally without installing CUDA, Python and C++ tools by hand.</p><p>For support, questions or installation issues, check the links above, the pinned comment, Patreon, Discord and the video comments, and send your logs.txt if a step fails. Thank you for watching.</p><h3>1-Click Windows and Linux (Ubuntu 22, 24, 26) AI Requirements Installers — Setup Your PC and Make It Ready for All Local AI Apps, Scripts, and More</h3><p><a href="https://www.patreon.com/SECourses/posts/1-click-windows-172290854">https://www.patreon.com/SECourses/posts/1-click-windows-172290854</a></p><p>Our AI Requirements Installers set up a Windows or Ubuntu PC for all our AI apps: one click on Windows, one command on Ubuntu. They install the latest NVIDIA driver, Python 3.10 to 3.14, CUDA 12.9 and 13.2, cuDNN 9.20, the C++ compilers, Git, Node.js, .NET and FFmpeg, set every path and variable, test everything for real and show the result in a clear report. Below you will see the real reports of a Windows 11 PC with an RTX 3090 Ti and a fresh Ubuntu 24.04 PC.</p><ul><li>This installer will get updated as needed like when we move to Torch 2.14 and CUDA 13.4.</li><li>I have been constantly getting asked how to make initial setup easy and here you have it.</li><li>Once you run this installers, your PC will be fully ready to run all of our AI apps with perfect speed, accuracy and quality.</li><li>For Windows, recommend version is Windows 11 but should work both Windows 10 and 11</li><li>For Ubuntu, recommended version is Ubuntu 24 but should work both Ubuntu 22 and 26 too</li><li>You can run multiple times with no issues if first install fails for any reason, don’t forget to restart system after running installation</li><li>This is only app that you need to run as administrator do not forget</li><li>All of my other apps, scripts, installers are run with double click or select and hit enter</li><li>Older manual installation : <a href="https://youtu.be/DrhUHnYfwC0">https://youtu.be/DrhUHnYfwC0</a></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zlgdlbQbyKSLmB9I.png" /></figure><h3>Why Our AI Requirements Installers</h3><ul><li>One click on Windows, one command on Ubuntu: right-click the .bat and choose Run as administrator, or run one sudo command in a terminal. Nothing has to be installed first.</li><li>Made for our apps and their full features: ComfyUI, SwarmUI, SECourses Musubi Trainer, AI Toolkit, Ultimate TTS, Premium Upscaler Pro, Video Captioner Pro and more, with torch.compile, Triton, flash-attention, sage-attention, xformers and insightface builds.</li><li>The latest NVIDIA driver: Windows gets the newest compatible WHQL driver for GeForce RTX 20-series and newer cards, checked with NVIDIA on every run. Ubuntu gets the newest driver Ubuntu offers, R595 or newer. The driver step runs last, and a working driver is never downgraded.</li><li>Several versions side by side: Python 3.10, 3.11, 3.12, 3.13 and 3.14 with 3.12 as the default, and CUDA 12.9 and 13.2 with 13.2 as the default. Ubuntu also gets the free-threaded Python 3.13t and 3.14t.</li><li>Tested for real: 78 tests on Windows and 66 on Ubuntu compile and start C++ and CUDA programs, build Python and CUDA extensions, encode a video and build a .NET app, in the same environment that a new terminal gets. On the RTX 3090 Ti, the CUDA test programs ran on the GPU with correct results.</li><li>Fast: the Windows 11 PC with the RTX 3090 Ti was installed, updated and tested in 20 minutes 45 seconds, with the CUDA and cuDNN installers already downloaded by an earlier run.</li><li>Made for PCs that already have tools: an existing Visual Studio is updated, a current Visual C++ runtime is left alone, and an older Python from the Python Install Manager is backed up while its virtual environments keep working.</li><li>A clear HTML report: every component, step, path, variable, test, download and backup on one page, opened in your browser at the end.</li><li>The right cuDNN every time: only cuDNN 9.20 for CUDA 13 is registered, so PyTorch always loads one matching cuDNN.</li><li>Ready after updates and in every terminal: compiling keeps working after Visual Studio updates on Windows, and on Ubuntu every new terminal is ready, bash, zsh and login shells included.</li><li>Safe on your PC: every change is backed up first and one command restores it. Ubuntu’s own Python stays as it is.</li><li>Run it again any time: finished parts are only checked, failed or interrupted parts are repaired, and components and the NVIDIA driver are updated.</li></ul><p>Here are both reports right after the installation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gRXVvKpXxI308RXO.png" /></figure><p>1 and 3: everything installed and verified, every test passed. 2: the full installation time on the Windows 11 PC with the RTX 3090 Ti. Its CUDA and cuDNN installers were already downloaded by an earlier run.</p><h3>What You Download</h3><p>You get small script files. They download everything else and check each download.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g2LN0D6STiOV11Sz.png" /></figure><p>1: on Windows you start the .bat. It comes in Windows_AI_Requirements_Installer.zip together with Windows_AI_Requirements_Installer.py, which must stay in the same folder. 2: on Ubuntu you run the .sh. The two text files explain every step and option.</p><h3>Windows AI Requirements Installer</h3><h3>How to Run It</h3><p>Extract the zip, right-click Windows_AI_Requirements_Installer.bat and choose Run as administrator. A normal double-click works too: it asks Windows for administrator rights by itself. The .bat brings its own temporary portable Python from <a href="https://python.org/">python.org</a>, so nothing has to be installed first.</p><p>When it finishes, the HTML report opens in your browser and logs.txt is saved next to the .bat. Open a new Command Prompt or PowerShell window to use the new PATH. Windows is never restarted for you: restart it when the report asks, and if a new NVIDIA driver asked for the restart, run the installer again afterwards to finish the tests.</p><h3>What It Installs</h3><p>The report starts with every installed component, its version and its folder:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*PXqmSY9ic4WxG-jS.png" /></figure><p>1: the latest NVIDIA WHQL driver, 617.42 on this PC. 2: Python 3.10 to 3.14 for all users in C:\Python310 to C:\Python314, and python, pip and py start Python 3.12. 3: CUDA 12.9 and 13.2 side by side with 13.2 as the default, and cuDNN 9.20. 4: FFmpeg 9 in a shared and a static build. 5: a 100 GB page file. You also get Visual Studio 2022 Build Tools with the MSVC C++ compilers and the Windows SDK, the latest Visual C++ runtime, Git, Node.js 24, the .NET 10 SDK for SwarmUI and Windows long paths.</p><h3>20 Steps, Every One Timed</h3><p>The Steps table shows each step with its status and time. Open the messages of any step to see exactly what it did:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SvSuRdufEQAusMom.png" /></figure><p>On this Windows 11 PC with an RTX 3090 Ti, CUDA 12.9 and 13.2 took 5 minutes, updating its Visual Studio Build Tools 3 minutes and the five Python versions 5 minutes. PATH and the variables were backed up before they changed, the NVIDIA driver step ran last, and the tests took 1 minute 48 seconds. The CUDA installers leave out their bundled older display driver, because the driver has its own step.</p><h3>The Latest NVIDIA Driver</h3><p>The installer keeps your NVIDIA driver up to date. It runs as the last installation step, right before the tests:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*G_nsJgN9HtLDpLHb.png" /></figure><p>1: the driver comes after CUDA, cuDNN and every other part, so a driver restart never comes before the other installers. 2: it finds GeForce RTX 20-series and newer desktop and laptop cards by their hardware IDs, even with no driver installed yet. 3: it asks NVIDIA’s live service for the newest compatible WHQL driver on every run, downloads the official NVIDIA-signed package and installs only the display driver. 4: a driver that is already the newest is kept, as here. Use — skip nvidia to keep your driver, or — only nvidia to update just the driver.</p><h3>Made for PCs That Already Have Tools</h3><p>This PC already had Visual Studio Build Tools, a .NET SDK and a Python 3.14 from the Python Install Manager. The installer worked with what was there:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZmK24S_9LUxP8E4e.png" /></figure><p>1: the existing Visual Studio Build Tools were updated to the latest 17.x. 2: the Visual C++ runtime was already current, so it was left alone. 3: the old Python 3.14 was unregistered with a backup, and py -3.14 now starts the new Python for all users. 4: a link keeps every existing virtual environment of the old Python working. The new .NET SDK was added next to the one already there.</p><h3>PATH in the Right Order</h3><p>The PATH order decides which Python, CUDA and cuDNN every app finds. The installer puts each folder at the top of the system and user PATH in the right order and shows the result in color:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KNdvCkacVmC0Li_H.png" /></figure><p>Green entries were added, blue ones moved to the top, and white ones are your own entries, kept as they were. cuDNN and CUDA 13.2 come first, then the Windows system folders, then Python 3.12 and the other versions, the C++ build paths and the FFmpeg shared build. 7: the entries of your other programs stay after ours. Your PATH before the change is listed below the table.</p><h3>Environment Variables</h3><p>The installer sets the variables that compilers and PyTorch extension builds look for, and it shows each one before and after:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Gk8QZQMyGMcpBtFb.png" /></figure><p>1: CUDA_HOME points to CUDA 13.2. 2: CUDNN_PATH points to the cuDNN 9.20 root that torch extension builds and CMake expect. 3: INCLUDE, LIB and LIBPATH make the C++ compiler work in every new terminal, like the Visual Studio developer prompt. 4: nvcc gets the flag it needs with the MSVC compiler. VSLANG keeps compiler messages in English for torch.compile, and values that were already correct stay unchanged.</p><h3>Visual Studio Updates Keep Compiling Working</h3><p>Visual Studio installs every compiler update into a new version folder. Our installer gives PATH, INCLUDE and LIB fixed paths that always lead to the current compiler and Windows SDK:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QxTVJihdMVhyxfvH.png" /></figure><p>1: a fixed path leads to today’s MSVC version. 2: a scheduled task points the paths at the new version at startup, at logon, after Visual Studio updates and every day, so open terminals keep working. 3: the 100 GB page file goes on the fastest suitable drive, chosen here from three drives.</p><h3>78 Tests, Run for Real</h3><p>At the end, the installer tests everything with exactly the PATH and variables that a new terminal or app gets:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*M0GXzVexx8-6MbQs.png" /></figure><p>All 78 tests passed. The first group checks that every command finds the right program: nvcc finds CUDA 13.2, cuDNN for CUDA 13 comes first, cl is the MSVC compiler, ffmpeg is the shared build, and the Windows system folders stay in front.</p><p>Every Python version is started and checked:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LeXCRbzRYhc_-Oty.png" /></figure><p>Each version runs with ssl, sqlite3, ctypes and tkinter, creates a virtual environment and runs pip in it, and has the headers and libraries that compiling extensions needs. The nvitop GPU monitor is installed too.</p><p>CUDA 12.9 and CUDA 13.2 are tested the way PyTorch uses them:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NdaRGPgKRAicnBxn.png" /></figure><p>The tests compile, link and start a CUDA program, and on the RTX 3090 Ti its GPU result was correct with both CUDA versions. They compile the Thrust and CUB code that PyTorch extensions use, compile GPU code for RTX 50xx (Blackwell) cards and compile a kernel at run time with NVRTC. cuDNN is built into a C++ program and loads the way onnxruntime and insightface load it.</p><p>The C++ tests build what torch.compile, Triton and pip build on your PC:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0nubY3gzJhuTjqQF.png" /></figure><p>An OpenMP library is built and loaded from Python, like torch.compile does on the CPU. A Python C extension is built the way pip builds insightface. CUDA 12.9 and CUDA 13.2 extensions are built with the compiler that PyTorch picks, the same way flash-attention and xformers are built. CMake and Ninja projects build too, and their CUDA programs ran on the GPU with correct results.</p><p>The last groups test FFmpeg, Git, Node.js, .NET and Windows itself:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5Kf34vglg-fXGvJx.png" /></figure><p>FFmpeg encodes a test video and reads it back, and its DLLs load the way torchcodec and torchaudio look for them. A .NET console app is created, built and run, as SwarmUI needs. A file path longer than 260 characters works, and the 100 GB page file is set.</p><h3>Checked Downloads and Backups</h3><p>Every download is checked before it is used, and every change is backed up first:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*PS-fnwKu7Q3fXk9z.png" /></figure><p>1: every installer must carry a valid digital signature from NVIDIA, Microsoft, the Python Software Foundation or its other publisher. 2: archives pass an integrity check. 3: send us logs.txt when you ask for help. 4: the backup folder holds HOW_TO_RESTORE.txt. 5: one command puts PATH and every variable back.</p><h3>Windows Options</h3><ul><li>— dry-run: shows the plan and changes nothing.</li><li>— only nvidia: checks for and installs just the latest compatible NVIDIA driver.</li><li>— skip nvidia: keeps your current driver during a full installation.</li><li>— only env,verify: runs only the steps you name. — skip skips steps, and — list-steps lists them.</li><li>— torch-test: installs PyTorch 2.13 with CUDA 13 in a temporary environment and runs real cuDNN convolution and BLAS tests on your NVIDIA GPU.</li><li>— restore-env “backup folder”: puts PATH and the variables back from a backup.</li><li>— no-browser: keeps the report from opening automatically.</li></ul><p>Pass them to the .bat from a terminal, for example Windows_AI_Requirements_Installer.bat — dry-run.</p><h3>Ubuntu AI Requirements Installer</h3><h3>How to Run It</h3><p>It works on Ubuntu 22.04, 24.04 and 26.04 LTS: desktop, server, cloud GPU machines and WSL2, and on systems based on them such as Linux Mint and Zorin OS (64-bit x86). Put Ubuntu_AI_Requirements_Installer.sh into a folder, open a terminal there and run:</p><p>sudo bash ./Ubuntu_AI_Requirements_Installer.sh</p><p>Enter your password, and the rest is automatic. At the end, the HTML report opens in your browser, and the report and logs.txt are saved next to the .sh file. Open a new terminal afterwards, and restart Ubuntu when the report asks for it after a new NVIDIA driver. The installation keeps running even if you close the terminal.</p><h3>What It Installs</h3><p>The Ubuntu report starts with the same overview:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Q3H0ljWDyy-hju-O.png" /></figure><p>1: Python 3.10 to 3.14 and the free-threaded 3.13t and 3.14t. python and py start Python 3.12, and Ubuntu’s own python3 stays as it is. 2: CUDA 12.9 and 13.2 with cuDNN 9.20 for CUDA 13 (CUDA 13.4 on Ubuntu 26.04). 3: FFmpeg 9 in a static and a shared build. 4: a swap file for 100 GB of swap in total. GCC, Clang, CMake, Ninja, Git with Git LFS, Node.js 24, the .NET 10 SDK and uv come too.</p><h3>17 Steps, the NVIDIA Driver Last</h3><p>Each step is listed with its time:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8uSulPX7UegxtN2E.png" /></figure><p>The installer first repairs the package system and adds the software sources for CUDA, Python and Node.js. On our fresh PC, the seven Python versions took 43 seconds and the tests 28 seconds. You also get aria2, 7-Zip, unrar, nvtop, htop and cloudflared for the apps’ share links.</p><p>On a PC with an NVIDIA GPU, the driver step installs the newest driver that Ubuntu offers for the card, R595 or newer (R615 on Ubuntu 26.04), with open kernel modules from Ubuntu’s signed packages. It upgrades the driver again on every run, never downgrades it, and gives RTX 50 (Blackwell) cards the open kernel module they need. It runs last, right before the tests.</p><h3>Ready in Every New Terminal</h3><p>The environment is written once for all users and loaded by every kind of terminal:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gyhxgCOdH7bGAOkE.png" /></figure><p>1: login shells, bash and zsh all load it. 2: use-cuda 12.9 switches one terminal to CUDA 12.9 when you need it. 3: CUDA_HOME, CUDA_PATH, CUDNN_PATH and the CUDA tools in PATH are set for all users. 4: this file is what every new terminal loads. The libraries are registered system wide, so no LD_LIBRARY_PATH is needed, and lines in ~/.bashrc and similar files that switch to another CUDA version are removed with a backup.</p><h3>66 Tests, Run as You</h3><p>Every test runs as your user with exactly the environment that a new terminal gets:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VoH8yff7wgtDPYqL.png" /></figure><p>All 66 tests passed. Each Python version runs with ssl, sqlite3, lzma, readline and tkinter, creates a virtual environment with pip and compiles a C extension.</p><p>The CUDA tests build and start programs with both CUDA versions:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*XK4PsQFxDVdgVlci.png" /></figure><p>They compile the Thrust and CUB code of PyTorch extensions and GPU code for RTX 50xx (Blackwell) cards, check that use-cuda 12.9 switches a terminal and back, and check that every program finds the CUDA and cuDNN libraries by name, like ONNX Runtime does.</p><p>Then come the C++ and FFmpeg tests:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HEz4Sa_IlwAysl0U.png" /></figure><p>GCC and Clang compile C++17, an OpenMP library is built and loaded like torch.compile does on the CPU, and a GPU launcher is built the way Triton compiles one at run time. FFmpeg encodes a test video and offers the x264, x265, SVT-AV1, VP9 and NVENC encoders. Its shared libraries are found the way TorchCodec looks for them, and the files for building PyAV and decord are in place.</p><p>The last groups check the tools and your own terminal:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gQPz5-axI4G8x7hE.png" /></figure><p>Git LFS, Node.js, .NET, aria2c, 7-Zip, unrar, jq, cloudflared, nvtop and htop all work, and a .NET app builds as SwarmUI needs. Your own terminal uses CUDA 13.2, and your ~/.bashrc, conda or pyenv settings keep the installed tools in reach.</p><h3>Before and After, Downloads and Backups</h3><p>The end of the report shows what changed and where everything is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yKfRuaxyffFf9e4b.png" /></figure><p>1: a fresh Ubuntu before, everything installed after. 2: every download is checked, the archives with SHA-256. 3: every APT package of the run is listed. 4: every changed file is backed up with HOW_TO_RESTORE.txt. 5: one command restores the files, and another removes what the installer added.</p><h3>Ubuntu Options</h3><ul><li>— yes: starts without a question, for scripts and cloud machines.</li><li>— dry-run: shows the plan without changing packages or settings.</li><li>— skip and — only: skip steps or run only some, for example — skip nvidia keeps your driver ( — list-steps lists them).</li><li>— swap-gb 32: sets the size of the swap file (0 = no swap file).</li><li>— torch-test: tests PyTorch 2.13 with CUDA 13 in a temporary environment: cuDNN convolution, BLAS, loaded libraries and torch.compile.</li><li>— repair: after a power loss or crash, checks every installed package and reinstalls damaged ones.</li><li>— restore-env “backup folder”: puts back the files it changed.</li><li>— uninstall: removes what the installer added outside of APT.</li></ul><p>Example: sudo bash ./Ubuntu_AI_Requirements_Installer.sh — yes — skip swap</p><h3>Run It Again Any Time</h3><p>Both installers are made to run again and again. Finished parts are only checked, failed or interrupted parts are retried and repaired, and components are updated to the versions in the installer. The NVIDIA driver is checked again on every run, too. Run the newest installer whenever you download it, and your PC stays ready for our newest apps.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZGp-_d5WJwnacXfeGifQIg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4cf820162dd5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Better Than ElevenLabs? Ultimate TTS Voice Training & Cloning in 600+ Languages]]></title>
            <link>https://medium.com/@furkangozukara/better-than-elevenlabs-ultimate-tts-voice-training-cloning-in-600-languages-035574499f79?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/035574499f79</guid>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 18:17:55 GMT</pubDate>
            <atom:updated>2026-10-06T18:17:55.012Z</atom:updated>
            <content:encoded><![CDATA[<p>Train your own AI voice with Ultimate TTS. This complete Windows tutorial covers voice cloning, automatic take selection, multilingual speech and OmniVoice full fine-tuning. The app also supports RunPod, SimplePod, Massed Compute and local Linux.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FivppvgjqAtg%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DivppvgjqAtg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FivppvgjqAtg%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/6929170d2c3eaf4b87ab721c620ddf81/href">https://medium.com/media/6929170d2c3eaf4b87ab721c620ddf81/href</a></iframe><h3>Links:</h3><ul><li><strong>Full YouTube Tutorial link : [ </strong><a href="https://www.youtube.com/watch?v=ivppvgjqAtg"><strong>https://www.youtube.com/watch?v=ivppvgjqAtg</strong></a><strong> ]</strong></li><li><strong>Download Ultimate TTS and read the source post: [ </strong><a href="https://www.patreon.com/SECourses/posts/ultimate-text-to-139297407"><strong>https://www.patreon.com/SECourses/posts/ultimate-text-to-139297407</strong></a><strong> ]</strong></li><li><strong>Windows Requirements tutorial: [ </strong><a href="https://youtu.be/DrhUHnYfwC0"><strong>https://youtu.be/DrhUHnYfwC0</strong></a><strong> ]</strong></li></ul><h3>Video Chapters:</h3><ul><li>00:00:00 My trained AI voice</li><li>00:00:38 600+ supported languages</li><li>00:01:19 Windows and cloud options</li><li>00:01:46 Your first voice clone</li><li>00:02:10 Generate and listen</li><li>00:03:18 Download ZIP and guides</li><li>00:03:46 Windows Requirements</li><li>00:04:28 Choose an install path</li><li>00:04:40 Extract the package</li><li>00:04:54 Run the installer</li><li>00:05:23 PyTorch, CUDA and wheels</li><li>00:05:54 Start the application</li><li>00:06:19 Models, voices, GPU presets</li><li>00:07:04 First load vs warm runs</li><li>00:07:46 Players and saved outputs</li><li>00:08:34 IndexTTS English example</li><li>00:09:02 AuK English example</li><li>00:09:52 Reference library and files</li><li>00:10:48 Select a reference range</li><li>00:11:38 Microphone recording</li><li>00:11:53 Reference text and transcription</li><li>00:13:19 Clone and Auto voice</li><li>00:14:38 Voice design tags</li><li>00:15:26 Language selection</li><li>00:15:40 Preview and Smart sentences</li><li>00:16:32 Section-length study</li><li>00:17:01 Sentence and section pauses</li><li>00:18:08 Pause tags</li><li>00:19:04 Check and set pronunciation</li><li>00:19:38 Save and reuse dictionary</li><li>00:21:10 Expressive tags</li><li>00:21:44 Keep which take</li><li>00:21:58 Fewest word errors</li><li>00:22:12 Rank takes and set checks</li><li>00:22:42 Ranking and fallback result</li><li>00:23:49 Whole-text candidates</li><li>00:24:36 60-line selection study</li><li>00:25:05 139-line Auto voice study</li><li>00:25:35 First vs selected take</li><li>00:26:42 Matching-language references</li><li>00:27:26 Arabic voice example</li><li>00:28:24 Chinese voice example</li><li>00:29:04 French voice example</li><li>00:29:46 German voice example</li><li>00:30:40 Italian voice example</li><li>00:31:22 Japanese voice example</li><li>00:32:03 Korean voice example</li><li>00:32:58 Polish voice example</li><li>00:33:38 Spanish voice example</li><li>00:33:49 Cross-language cloning</li><li>00:35:48 Speaking-rate comparison</li><li>00:36:49 Target duration modes</li><li>00:37:39 Best quality: 32 steps</li><li>00:38:25 Fast: 16 steps</li><li>00:38:54 Seeds and advanced sampling</li><li>00:39:44 WAV, MP3 and filenames</li><li>00:40:01 Reference and subtitle export</li><li>00:40:16 Inspect the exported files</li><li>00:41:10 Recent outputs as references</li><li>00:41:52 Audio tuning comparison</li><li>00:43:09 Tuning and edge silence</li><li>00:43:30 SRT import and cue timing</li><li>00:44:22 Create a narrated MP4</li><li>00:45:08 Batch inputs and filenames</li><li>00:45:43 Shared and per-file voices</li><li>00:46:07 Batch execution and progress</li><li>00:46:47 Inspect batch results</li><li>00:48:00 Save a personal preset</li><li>00:48:18 Load settings and reference</li><li>00:49:26 Recent settings and theme</li><li>00:49:45 Select GPU and memory tier</li><li>00:50:02 BF16 and ConvRot INT8</li><li>00:50:16 Measured memory use</li><li>00:50:30 Measured synthesis time</li><li>00:50:44 Compare both variants</li><li>00:52:25 Section batch size</li><li>00:52:39 Load models and find files</li><li>00:53:03 VRAM benchmark and tier cap</li><li>00:53:56 Train a new OmniVoice</li><li>00:54:15 Source recordings and text</li><li>00:54:44 Reserve evaluation sessions</li><li>00:55:00 Dataset folder and scan</li><li>00:55:37 Transcript and alignment</li><li>00:55:57 Segment duration</li><li>00:56:13 Padding and caption cleanup</li><li>00:56:28 Quality filters</li><li>00:56:47 Prepare and inspect clips</li><li>00:58:02 Speaker and transcript audit</li><li>00:58:50 Review flagged clips</li><li>00:59:21 Correct words and boundaries</li><li>01:00:20 Build the OmniVoice cache</li><li>01:00:47 Dataset and training name</li><li>01:01:06 Full fine-tuning</li><li>01:01:33 Full, LoRA and DoRA</li><li>01:01:50 Held-out voice comparison</li><li>01:02:05 Epochs and Training plan</li><li>01:02:22 Token budget and accumulation</li><li>01:02:39 Learning rate and memory</li><li>01:03:14 Checkpoints and train state</li><li>01:03:30 Continue vs weights-only</li><li>01:03:47 Evaluation and final test</li><li>01:04:18 Fluency analysis and filters</li><li>01:05:16 Start training</li><li>01:05:49 Speed, loss and validation</li><li>01:06:18 Hear a training sample</li><li>01:06:45 Stop and resume controls</li><li>01:07:05 800-update completion</li><li>01:07:39 Checkpoint analysis</li><li>01:07:55 Automatic speech comparison</li><li>01:08:15 Reference audition</li><li>01:08:37 Speaking-rate calibration</li><li>01:08:53 Save an INT8 voice copy</li><li>01:09:46 Load the trained preset</li><li>01:10:14 Generate held-out speech</li><li>01:11:12 Base vs trained voice</li><li>01:12:15 Configure Checkpoint Grid</li><li>01:13:24 Hear the six grid cells</li><li>01:14:34 Trained Auto voice</li><li>01:15:22 Trained INT8 voice</li><li>01:16:13 Updates, backup and help</li><li>01:17:01 Your next voice</li></ul><p>This video is for anyone who wants a complete voice workflow, from a short reference to a reusable trained voice. Start with the included reference and save your first working preset.</p><p>For support, questions, feature requests, installation issues and updates, use the download post, pinned comment, Patreon, Discord and video comments. Thank you for watching.</p><p>Ultimate Text To Speech Generator With Voice Cloning turns your scripts, subtitles and whole chapters into natural speech on your own PC, with three speech models in one app: IndexTTS 2.5, OmniVoice and AuK. Clone any voice from a short clip, design a new voice from a few words, edit recordings, and train your own voice from your videos. Below you will see every feature with real screenshots from the app.</p><p>OmniVoice DoRA (better version of LoRA) trainings is yielding literally better than ElevenLabs with our special pipeline + Keep which take inference system.</p><h3>Why Ultimate Text To Speech Generator</h3><ul><li>Three speech models in one app: IndexTTS 2.5, OmniVoice and AuK. Switch them in the header; each keeps its own settings.</li><li>Clone any voice in seconds: drop an audio or video file and type your text. On an RTX 5090, IndexTTS 2.5 made 31 seconds of speech in 18 seconds from a 14.6 second part of a video.</li><li>Take quality: the app renders several takes of every section, Whisper checks them, and the take that sounds most like the voice without a word error is kept. In the developer’s test on 60 lines, word errors fell from 0.65 % to 0.20 % with OmniVoice.</li><li>New voices without a recording: describe a voice in one sentence with AuK, or pick voice tags with OmniVoice.</li><li>AuK Audio Editing: 23 editing tasks: replace, add or remove words, change speed, pitch, volume or emotion, clean up noise and separate voices. Replacing two words in a 10 second clip took 1.7 seconds.</li><li>Train your own voice for any model: from your videos and subtitles to a finished voice with a ready-to-use preset. Four recordings became 205 clean training clips in 50 seconds.</li><li>Made for long narration: Smart sentences, exact pause tags, a live section preview and a pronunciation dictionary.</li><li>Subtitles to speech and MP4: 10 subtitle formats, speech timed to every cue, and an MP4 with your image in the same run.</li><li>Batch generation: whole folders of scripts and subtitles in one click.</li><li>Ready for your GPU: 7 GPU presets from 6 GB to 32 GB, BF16 or INT8 ConvRot models, block swap and a built-in VRAM benchmark.</li><li>1-click installers: Windows, RunPod, SimplePod, Massed Compute and Linux, with PyTorch 2.14.1 and CUDA 13. The models download automatically, and every model weight file is checked with SHA256.</li><li>Frequent updates: four releases between 2 and 4 October 2026, with a built-in Help tab and changelog.</li></ul><p>Here is the app right after a voice clone. The voice came from a 14.6 second part of one of our tutorial videos, and 31 seconds of speech were ready in 18 seconds.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Z5Kcwv9mIcFVr9gt.png" /></figure><p>1: choose IndexTTS 2.5, OmniVoice or AuK in the header. 2: the preset for your GPU is selected on the first start. 3: any audio or video file can be the voice. 4: type your script. 5: the result plays right away, with the elapsed time, speed and VRAM next to it.</p><h3>1-Click Installation</h3><h3>What You Download</h3><p>You get one small zip file with the installers for every platform.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*olryyokkDOLPMNTk.png" /></figure><p>Extract it into a short folder path without spaces, double-click Windows_Install_or_Update.bat, then start the app with Windows_Start_App.bat. Run the same installer again at any time to update.</p><h3>Latest PyTorch and CUDA 13</h3><p>The installer makes its own Python 3.12 virtual environment, so your other apps stay untouched. It installs PyTorch 2.14.1 with CUDA 13, Transformers, Triton for Windows and every other package with uv, a very fast package installer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZKADDgKrdLzw1NC0.png" /></figure><p>PyTorch, torchvision and torchaudio always come from the same CUDA 13 build. Our fresh install took 5 minutes 39 seconds on our PC, models included.</p><h3>Automatic Model Downloads</h3><p>At the end, the installer downloads the IndexTTS 2.5 models and their helper models by itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*z8y8jtKyuYsVmg95.png" /></figure><p>Large files download over 16 connections, and every model weight file is checked with SHA256, so a broken download is caught right away. Running the installer again skips the large files you already have.</p><h3>Start the App</h3><p>Double-click Windows_Start_App.bat. On the first start, the app reads your GPU memory and selects the matching GPU preset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mt_-xBpoySQGr4Er.png" /></figure><p>The app was ready in 9.5 seconds on the first start and in 4 to 6 seconds later. OmniVoice, AuK and the built-in Whisper download automatically the first time you use them.</p><h3>Cloud GPUs: RunPod, SimplePod and Massed Compute</h3><p>You can also run the app on a cloud GPU. The cloud installers set up everything with one command and download the models for you. A Gradio share link lets you use the app from any device.</p><ul><li>SimplePod: <a href="https://simplepod.ai/ref?user=secourses">register here</a> and use <a href="https://dash.simplepod.ai/account/explore/100/ref-secourses/">this template</a>.</li><li>RunPod: <a href="https://get.runpod.io/955rkuppqv4h">register here</a> and use <a href="https://get.runpod.io/SECourses_CU13">this template</a>.</li><li>Massed Compute: <a href="https://vm.massedcompute.com/signup?linkId=lp_034338&amp;sourceId=secourses&amp;tenantId=massed-compute">register here</a> and use our coupon SECourses.</li></ul><p>The step-by-step commands are in the instruction files inside the zip.</p><h3>Requirements</h3><p>On Windows you need Python 3.12, Git, FFmpeg, CUDA 13, cuDNN 9.17 and Visual Studio with the C++ tools. <a href="https://youtu.be/DrhUHnYfwC0">This tutorial</a> shows every step, and the same setup runs all our AI apps.</p><h3>Three Speech Models in One App</h3><p>Every model has its own strengths, and you switch between them with one click in the header.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WTcx7XkHiixAC-qm.png" /></figure><p>IndexTTS 2.5 gives you emotion control and reads five languages. OmniVoice clones voices, designs voices from tags and speaks many languages. AuK clones voices, designs voices from a plain sentence and edits recordings in its own tab. Each model keeps its own settings, and every preset stores all three.</p><h3>GPU Presets for 6 GB to 32 GB Cards</h3><p>Seven read-only presets fit the whole app to your card: generation, dataset preparation and training together. On the first start, the app selects the preset of your GPU.</p><p>Where to find it: the Universal preset list sits at the top of the page, above every tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FigP9Uv3NmiW-5RZ.png" /></figure><p>Open the Universal preset list to see every preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pweRKK1P0Y237nZb.png" /></figure><p>Change anything you like, type a name and click Save to keep your own preset. It stores every setting of every tab and of all three models. Every training adds a ready-to-use preset of its new voice to this list. The Help tab shows what each GPU preset uses:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tzSQeG01wL_21rRh.png" /></figure><p>A smaller card pays with speed first: helper models move to the CPU, and the 6 GB preset streams 22 of the 24 GPT blocks from RAM.</p><h3>Voice Cloning From Any Audio or Video</h3><p>A clean clip of a few seconds is enough to clone a voice. Audio and video files both work, and the preview plays exactly what the app will use.</p><p>Where to find it: open the Voice Generation tab. Reference Voice is the left column.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OEcLARv6iEFjw0jf.png" /></figure><p>Here is the Reference Voice column up close:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*J28pKxMpJDZwM7WQ.png" /></figure><p>Drop a file, record from your microphone, or load any file on your PC by its path. Time ranges such as 0.2:14.8 keep only the clean part of a long recording, and several ranges are joined in order.</p><h3>Text, Smart Sentences and Pauses</h3><p>Long text is split into sections before the voice reads it. Smart sentences packs whole sentences into each section.</p><p>Where to find it: in the Voice Generation tab, Text &amp; Timing is the middle column.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9ZgpMFeTTAOVxSry.png" /></figure><p>Here are the text controls:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uXZfxKyhgn5vVCpg.png" /></figure><p>Tags such as [pause:800ms] insert an exact silence. With a trained voice, the line length and the pauses between sentences come from the speaker’s own recordings. Below the text, the live section preview shows every section before you generate:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*S-ex-gIx0RJYIhbv.png" /></figure><p>You see the tokens and words of each section and every pause, so you can fix the text before you spend any GPU time.</p><h3>Pronunciation Check and Dictionary</h3><p>Names and technical words are the hardest part of any narration. The pronunciation check finds the words the model may not know and suggests a reading for each.</p><p>Where to find it: in the Voice Generation tab, click Pronunciation check &amp; dictionary under the live section preview.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OZTYep9rDjshZx7z.png" /></figure><p>Click Check unknown words:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qW55KeIe6urMdDDy.png" /></figure><p>Readings come from the CMU dictionary, acronym and CamelCase splitting, and letter rules. Save the good ones to your dictionary, and the app applies them in Voice Generation, Batch Generation and the live preview, for IndexTTS 2.5 and OmniVoice.</p><h3>Take Quality: The Best Take of Every Section</h3><p>Take quality picks the best take of every section. It renders several takes, lets the built-in Whisper check every word, and keeps the take that sounds most like the voice without a word error. It works for all three models, and the GPU presets switch it on.</p><p>Where to find it: in the Voice Generation tab, Take quality is in the right column, under the live log.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uF23RxU9CK-J8ebu.png" /></figure><p>Here is Take quality after an OmniVoice clone of the same video clip:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IbR4GtMW716dChdD.png" /></figure><p>OmniVoice rendered 10 takes of our 32 second script in 26.5 seconds, and the app kept the most similar take with 0 word errors. In the developer’s test on 60 lines of voice cloning, word errors fell from 0.65 % to 0.20 % with OmniVoice and from 0.91 % to 0.35 % with IndexTTS 2.5.</p><h3>OmniVoice: Clone, Design or Auto Voice</h3><p>OmniVoice clones a reference voice, designs a new voice from tags, or picks a voice by itself.</p><p>Where to find it: choose OmniVoice in the header. Its section opens in the Voice Generation tab, under Voice LoRA / DoRA.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*y9an4CMYn45vGUvj.png" /></figure><p>Here is the OmniVoice section with a designed voice:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6LX432vp8V-Fu4_n.png" /></figure><p>We picked four tags: female, young adult, moderate pitch and british accent. The new voice spoke 6.1 seconds of speech in 1.2 seconds, with no reference recording at all.</p><h3>AuK: Describe a Voice in Words</h3><p>AuK clones voices, designs a voice from one plain sentence, and lets a trained AuK voice speak without a reference.</p><p>Where to find it: choose AuK in the header. Its section opens in the Voice Generation tab, under Voice LoRA / DoRA.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AiIGS4vEJoFVvhqw.png" /></figure><p>Here is the AuK section with a described voice:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uf6MW53DqUxEadHc.png" /></figure><p>“A calm middle-aged man with a deep, warm voice, speaking slowly and clearly, like a documentary narrator.” AuK made 6.6 seconds of speech from that sentence in 1.5 seconds. AuK speaks English and Chinese.</p><h3>Emotion Control</h3><p>IndexTTS 2.5 separates the voice from the emotion, and the app gives you four ways to set the emotion.</p><p>Where to find it: with IndexTTS 2.5 selected, scroll down in the Voice Generation tab and click Emotion Control. The Open / close all sections button at the top opens it too.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TLN5XL8q1lo0a67K.png" /></figure><p>Here is Emotion Control with the emotion sliders selected:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GoN6jtTE1fekXxuH.png" /></figure><p>Keep the speaker’s own tone, copy the delivery of an emotion clip, mix joy, anger, sadness, fear, disgust, depression, surprise and calm, or describe the emotion in words. Emotion weight sets how strong it is.</p><h3>Timing, Output Formats and Word Timestamps</h3><p>Make the speech fit your video, save it in the format you need, and get subtitles of every word.</p><p>Where to find it: in the Voice Generation tab, click Segmentation &amp; Timing, Output and Execution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uJRUu1UQpSiB8u5J.png" /></figure><p>The three sections opened:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_iG8zQUi9UtYKZLH.png" /></figure><p>Ask for an exact length, such as 25 seconds, and Natural mode paces the speech to fit. Every run saves a WAV file, one tick adds an MP3 up to 320k, and another adds SRT and VTT subtitles with the timing of every word. Isolated subprocess mode frees all VRAM after each job.</p><h3>Subtitles to Timed Speech and MP4</h3><p>Turn any subtitle file into speech that follows its timing, ready for dubbing, translations and slide videos.</p><p>Where to find it: in the Voice Generation tab, the Captions and Still image for MP4 boxes are under the text controls.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MYf_AtlxrahCc_2f.png" /></figure><p>Drop your subtitle file and tick Use caption cue timing:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7Gv0rPQSUm_3jijI.png" /></figure><p>SRT, VTT, SBV, ASS/SSA, SUB, LRC, TTML/DFXP, SAMI, JSON and TSV all work, and the app recognizes the format from the file itself. Add an image, and the same run also makes an MP4:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YRld8QqTSvwuCZ2g.png" /></figure><p>Our 6-cue subtitle file became exactly 30.00 seconds of speech, with every cue starting on time.</p><h3>AuK Audio Editing</h3><p>AuK edits real recordings: change the words, the delivery or the sound, and separate voices.</p><p>Where to find it: choose AuK in the header, then click the AuK Audio Editing tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vN_ejyc-zUTUmJ7N.png" /></figure><p>Here is a word replacement:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MbLKnsWS_cLY7_SN.png" /></figure><p>We replaced “quick test” with “short demo” in a 9.9 second recording. The edit took 1.7 seconds, and Whisper heard the new words in the result. The 23 tasks also insert or remove words, change speed, pitch, volume and emotion, turn speech into a whisper, add a laugh, clean up noise and reverberation, and keep one speaker or the voices from music.</p><h3>Batch Generation</h3><p>Generate many scripts or subtitle files in one run.</p><p>Where to find it: click the Batch Generation tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MEqiSN_-QzFVP9TN.png" /></figure><p>Drop your files, paste your texts or point the app at a folder, then click Generate batch:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RcYAVPTUSN7AvTPF.png" /></figure><p>Our three stories became 116 seconds of speech in 65 seconds with OmniVoice, every section picked from 10 takes. Use one voice for all files, or a matching audio file next to each script for a different voice per file. Continue after item errors keeps the batch going.</p><h3>Train Your Own Voice</h3><p>A trained voice follows the speaker much more closely than a single reference clip: the pace, the pauses and the way each word is said. You can train IndexTTS 2.5, OmniVoice and AuK, and the app takes you from videos and subtitles to a finished voice.</p><h3>Prepare the Dataset</h3><p>Point the app at your videos or audio files with their subtitles. It cuts them into clean clips of whole sentences.</p><p>Where to find it: click the LoRA Dataset Preparation tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AhYw8ml235OMYzZH.png" /></figure><p>Here is a real run on four of our tutorial recordings:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Zy5QCQK4NBrpRiIN.png" /></figure><p>Your subtitles give the words and Whisper gives the exact word timing. The four recordings, 46.8 minutes in total, became 205 clean clips of 40.6 minutes in 50 seconds. Every clip can be checked before training:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TgoXFUQuDxe27WUk.png" /></figure><p>The clips are 4 to 16 seconds long, and every clip keeps its text.</p><h3>Voice and Transcript Audit</h3><p>The audit checks every clip against a clean recording of the speaker and against its text, and builds a separate training dataset from the clips that pass.</p><p>Where to find it: in the LoRA Dataset Preparation tab, click Voice and transcript audit at the bottom.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*a9rcKZtn3LoN6SF4.png" /></figure><p>Here is the audit of our 205 clips:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TXnRpCTWgDcfA1dB.png" /></figure><p>It checked 205 clips in 1 minute 14 seconds and kept 194. A second Whisper model rescued 7 good clips that the first one misheard. A whole recording can be held out, so the app tests the voice on speech it never trained on.</p><h3>Start the Training</h3><p>Choose your dataset, give the voice a name and click Start training. The defaults are the settings we measured on real voices.</p><p>Where to find it: click the Voice Training tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OhD4CtG8AgucBTYE.png" /></figure><p>Here is the tab with our audited dataset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hcLQfVAVET2gkrPC.png" /></figure><p>The GPU VRAM preset fits the training to your card. DoRA and LoRA train small adapters; full fine-tuning trains the speech model’s own weights on 16 GB cards and larger (AuK 32 GB). The fluency filter can train on fluent clips only. During training the dashboard updates live:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NRQgp64ZDvEcwvLM.png" /></figure><p>Our IndexTTS 2.5 voice trained at about 5 to 6 steps per second with 2.3 GB of VRAM. The app checks the voice on held-out clips every epoch and keeps the best checkpoint.</p><h3>Automatic Checks After Training</h3><p>When training ends, the app keeps working for you. It renders the same sentences with the base model and with the checkpoints, then compares them with the speaker’s real recordings of those sentences. It also calibrates the speaking rate, trains a voice decoder adapter, tunes the decoding settings and auditions reference clips. Here is the comparison for our voice:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qcDmxY7cERZb3YG5.png" /></figure><p>Likeness to the speaker’s real recordings rose from 0.810 with the base model to 0.867 for the chosen checkpoint with its voice decoder adapter, and the word error rate fell from 2.4 % to 2.3 %. The app picked that checkpoint by itself and set it as the voice’s default.</p><h3>A Ready-to-Use Preset</h3><p>Every training ends with a ready-to-use preset of the new voice: our training saved my_voice_Takes_5 in the Universal preset list you saw above. It loads the chosen checkpoint, the best reference clip from the audition, the calibrated speaking rate of 0.86, the decoding settings from the sweep, and Take quality with up to 5 takes per section.</p><h3>Use Your Trained Voice</h3><p>Select your voice in Voice Generation, and the app loads everything it learned.</p><p>Where to find it: in the Voice Generation tab, the Voice LoRA / DoRA panel is right under the three columns. You can also load the voice’s ready-made preset from the Universal preset list.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tKBgEJpLkuNSF_nd.png" /></figure><p>Here is the panel after loading the preset of our new voice:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r_MlQoLVC1QIfSuk.png" /></figure><p>The panel shows the voice’s calibrated pace, the line length and the pauses measured from its own training clips, and the decoding its sweep adopted, and applies them to your text. With this preset our voice read the 88-word script as 35.2 seconds of speech in 28 seconds, and the built-in Whisper found 0 word errors in every section.</p><h3>Compare Checkpoints by Ear</h3><p>The Checkpoint Grid tab renders the same texts with every checkpoint you choose, so you can listen to them side by side.</p><p>Where to find it: click the Checkpoint Grid tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tUkL_QMhB5Eo14Oo.png" /></figure><p>Here is a grid of our voice with the base model and three checkpoints:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QSRR36UHRxtGIhnG.png" /></figure><p>Every cell uses the same texts, reference and seed, so you hear only the difference between the checkpoints. Every grid is saved and opens again at any time, and Calibrate speaking rate from this grid measures the voice’s real pace.</p><h3>Models and Performance</h3><h3>Runtime</h3><p>Fit the models to your GPU in one place.</p><p>Where to find it: click the Models &amp; Performance tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qsRUAoYiMfERG7z2.png" /></figure><p>Here are the runtime controls:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fE7XGLywR76GrCLA.png" /></figure><p>Choose a VRAM tier, and the fit check shows the expected peak memory before anything loads. Pick the BF16 model or the INT8 ConvRot model, which saves memory. Unload model / free VRAM clears the GPU in one click.</p><h3>Block Swap and Model Placement</h3><p>Small cards can stream parts of the model from system RAM.</p><p>Where to find it: in the Models &amp; Performance tab, click Block Swap &amp; Memory and Auxiliary Model Residency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zEJZFu64JSscFFLx.png" /></figure><p>Here is the 6 GB tier:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YoyvdUSP2wcZ2iTB.png" /></figure><p>The 6 GB tier streams 22 of the 24 GPT blocks from RAM and keeps the reference models on the CPU. Every helper model can stay on the GPU, stay on the CPU or load on demand.</p><h3>Model Files and VRAM Benchmark</h3><p>Check your model files and test a GPU tier before a long job.</p><p>Where to find it: in the Models &amp; Performance tab, click Model Files &amp; Downloads and VRAM Benchmark.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lH2YjuWsO6zFUqsO.png" /></figure><p>Both sections after a real test:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nZIkSyO_aofkw4dA.png" /></figure><p>The INT8 ConvRot GPT is 1.1 GB instead of 3.1 GB and downloads from inside the app. The VRAM benchmark ran the 8 GB tier with its memory limit in place on our card: it fit with a 3.5 GB peak and made 25.9 seconds of speech in 19.3 seconds.</p><h3>Built-in Help and Changelog</h3><p>You never need to leave the app to learn it.</p><p>Where to find it: click the Help tab or the Changelog tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xwu8ydDubPOS7QPQ.png" /></figure><p>The Help tab starts with a four-step quick start:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*GUijQEi2-1HMmGwx.png" /></figure><p>It explains every model, every workflow and every parameter. The Changelog tab lists every release in plain words:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7s4NQAgj-1bNEsA7.png" /></figure><p>The newest release is always at the top.</p><h3>Get Ultimate Text To Speech Generator</h3><p>Download the latest zip file attached to this post, extract it and run the installer for your platform. To update, get the newest zip, overwrite the old files and run Windows_Install_or_Update.bat again.</p><ul><li>Requirements tutorial: <a href="https://youtu.be/DrhUHnYfwC0">Python, Git, FFmpeg, CUDA and C++ tools step by step</a>.</li><li>Support: <a href="https://discord.com/channels/772774097734074388/1079506787734134844">our Discord channel</a>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l_x8D1A3r_KHjXURyhaBmQ.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=035574499f79" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Whisper-WebUI Premium Transcribes a 10-Minute Video in 4 Seconds: Full Windows Guide]]></title>
            <link>https://medium.com/@furkangozukara/whisper-webui-premium-transcribes-a-10-minute-video-in-4-seconds-full-windows-guide-0d1740d6b59a?source=rss-fb0739c3d821------2</link>
            <guid isPermaLink="false">https://medium.com/p/0d1740d6b59a</guid>
            <dc:creator><![CDATA[Furkan Gözükara - PhD Computer Engineer, SECourses]]></dc:creator>
            <pubDate>Thu, 01 Oct 2026 22:35:05 GMT</pubDate>
            <atom:updated>2026-10-01T22:35:05.330Z</atom:updated>
            <content:encoded><![CDATA[<h3>App download link &gt; <a href="https://www.patreon.com/SECourses/posts/whisper-webui-to-145395299">https://www.patreon.com/SECourses/posts/whisper-webui-to-145395299</a></h3><h3>Full tutorial link &gt; <a href="https://www.youtube.com/watch?v=yzNhkC3aQuA">https://www.youtube.com/watch?v=yzNhkC3aQuA</a></h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FyzNhkC3aQuA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DyzNhkC3aQuA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FyzNhkC3aQuA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/8e6183734b027e29dc2a10feed625713/href">https://medium.com/media/8e6183734b027e29dc2a10feed625713/href</a></iframe><p>Whisper-WebUI Premium turns video and audio into subtitles, searchable transcripts and timed data on your own Windows PC. See a real 10-minute recording transcribed in 4 seconds, then follow the complete fresh installation and every main workflow. This tutorial covers all 3 engines, researched quality presets, automatic model downloads, 6 export formats, custom vocabulary, folder batches, YouTube sources, microphone recording, speaker labels, subtitle translation and voice/music separation.</p><p>Links:</p><p>Download Whisper-WebUI Premium and read the full feature guide: [ <a href="https://www.patreon.com/SECourses/posts/whisper-webui-to-145395299">https://www.patreon.com/SECourses/posts/whisper-webui-to-145395299</a> ]</p><p>Windows Requirements tutorial: [ <a href="https://youtu.be/DrhUHnYfwC0">https://youtu.be/DrhUHnYfwC0</a> ]</p><p>You will learn how to download the latest ZIP, locate it in Chrome, move it to a permanent folder and extract the whole package. Read the Windows Requirements section and follow its linked setup tutorial first if needed. Use an English-character path without spaces, such as G:\whisper_tutorial. Run the Windows installer and app starter by double-clicking their BAT files or selecting them and pressing Enter; never run them as administrator. Keep the start window open while using the local browser interface. We then load a video by upload or full file path, choose a language and output formats, generate the first subtitles, and inspect the saved files and transcript. Use SRT for conventional video subtitles, WebVTT for compatible web players, and TXT for searching or rewriting the transcript. LRC, JSON and TSV keep timed information for other tools. Select several formats in the same run and inspect matching filenames in the outputs folder.</p><p>Main topics include multilingual Whisper and optimized models, Canary-Qwen for English speech, Insanely Fast Whisper, GPU-aware presets, INT8 ConvRot, PyTorch 2.13 and CUDA 13 precompiled packages. Save your own preset with vocabulary, reload its settings, reset engine defaults or delete a user preset. Inspect compute precision, beam search, prompts and hotwords, repetition and silence controls, batch size, model offloading, word timestamps and subtitle punctuation. See a long lecture result and stop a separate worker with Cancel Generation. The filter demonstrations cover background music removal, Silero voice detection and diarization for a two-man conversation. Batch Processing scans nested input folders, mirrors their structure in a separate output folder, skips existing results and offers deliberate overwrite. The YouTube tab reads a video link or processes recent videos from a channel. Live Mic previews incoming text; Record Then Generate creates subtitles after recording. Local NLLB translates an English SRT to Spanish while preserving all 10 cue timings and speaker labels; the DeepL API controls provide another route with your own key. BGM Separation saves individual voice and instrumental tracks. SRT, WebVTT, TXT, LRC, JSON and TSV can be selected together and downloaded in one ZIP. Finally, switch themes, open or close the settings panels, and follow the package update route.</p><p>Use the chapters to follow the fresh Windows install or jump straight to engines, presets, batch folders, speaker labels, microphone modes, translation and audio separation.</p><p>Chapters:</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=0">00:00:00</a> A ten-minute video subtitled in four seconds</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=69">00:01:09</a> See the transcript and choose the right exports</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=108">00:01:48</a> Download the ZIP and check Windows requirements</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=163">00:02:43</a> Move and extract the complete installer package</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=212">00:03:32</a> Fresh Windows installation with the BAT installer</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=286">00:04:46</a> Launch the app and make your first subtitles</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=430">00:07:10</a> Three engines and the long lecture result</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=560">00:09:20</a> Save and reload your own working preset</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=634">00:10:34</a> Useful advanced controls without rebuilding your setup</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=924">00:15:24</a> Music removal, voice detection and speaker labels</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=1055">00:17:35</a> Transcribe nested folders and reuse finished outputs</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=1150">00:19:10</a> Video links and whole-channel processing</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=1216">00:20:16</a> Live preview and record-then-generate microphone modes</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=1313">00:21:53</a> Translate subtitles and preserve their timings</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=1390">00:23:10</a> Save separate voice and instrumental audio tracks</p><p><a href="https://youtu.be/yzNhkC3aQuA?t=1439">00:23:59</a> Updates, cloud support and the complete workflow</p><p>The opening result is a 10-minute 56-second recording completed in 4 seconds on an RTX 5090, using the Canary Qwen Best Quality preset, batch size 16 and the optimized Canary INT8 ConvRot model with the model and caches ready. The app reports the elapsed time and result on screen.</p><p>The package also supports RunPod, SimplePod and Massed Compute for cloud GPU workflows. This video teaches Windows installation and use; the included platform instructions cover those cloud routes.</p><p>Read the linked post for release notes, illustrated feature explanations and current downloads. The app link is also in the pinned comment. For questions, installation support, feature requests and feedback, use the post and video comments.</p><p>Whisper-WebUI Premium turns any video or audio into accurate subtitles and transcripts on your own PC. On an RTX 5090, our own INT8 ConvRot engine transcribed a 1 hour 29 minute lecture in 32 seconds. Every setting comes ready with researched best-quality presets. Below you will see every feature with real screenshots from the app.</p><h3>Why Whisper-WebUI Premium</h3><ul><li>Our own INT8 ConvRot engines: Whisper large-v3 runs 3.4x faster and NVIDIA Canary-Qwen 2.5B runs 8.9x faster than the standard models, on the same video and the same GPU.</li><li>Same accuracy as the original models: measured through the app on 14 public English test sets, 48 hours of audio.</li><li>Real speed: a 10 minute video in 9 seconds, a 1.5 hour lecture in 32 seconds.</li><li>3 engines in one app: Whisper, Insanely Fast Whisper and NVIDIA Canary-Qwen 2.5B, with 21 Whisper models and 100 languages.</li><li>Ready presets: researched best-quality settings for every engine, plus your own saved presets.</li><li>6 output formats in one run: SRT, WebVTT, TXT, LRC, JSON and TSV, with a one-click ZIP download.</li><li>Everything in one place: batch folders, YouTube links and whole channels, live microphone, speaker labels, background music remover, voice detection filter and subtitle translation to 200 languages.</li><li>1-click installers: Windows, RunPod, SimplePod, Massed Compute and Linux, with PyTorch 2.13, CUDA 13 and precompiled Flash Attention, xFormers, SageAttention and Triton.</li><li>Automatic model downloads with live progress, frequent updates and support.</li></ul><p>Here is the app right after a job. A 10 minute 54 second video was transcribed in 9 seconds, 70 times faster than real time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Vub4cypqI2G8CZUz.png" /></figure><p>1: pick a ready preset. 2: download all subtitle files in one ZIP. 3: preview your video instantly. 4: watch the transcription live. 5: choose from 3 engines and our INT8 ConvRot models.</p><h3>Speed: Our INT8 ConvRot Engines</h3><p>We built INT8 ConvRot versions of Whisper large-v3, Whisper large-v1 and NVIDIA Canary-Qwen 2.5B, and our own GPU engine to run them. The model files are smaller too: 1.6 GB instead of 3.1 GB for Whisper, and 2.9 GB instead of 5.1 GB for Canary-Qwen. Here is the same video on the same RTX 5090 with the same settings:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lHasyBOfgKNU6j4N.png" /></figure><p>Long files are just as fast. Our 1 hour 29 minute lecture took 1 minute 15 seconds with our INT8 Whisper large-v3 (71x real time) and only 32 seconds with Canary-Qwen (168x real time). This is the Canary-Qwen run in CMD:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UKtUyjkY96M6RMWU.png" /></figure><p>After each job the model waits in RAM and your VRAM is free. The next job moves it back to the GPU in under a second, so it starts right away.</p><h3>Same Accuracy as the Original Models</h3><p>Speed only matters with accuracy. We measured every model through the app on 14 public English test sets with human transcripts: 2,700 short clips and 120 long recordings, 48 hours in total.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ca2LKR-oIGTbIMeK.png" /></figure><p>On the Open ASR Leaderboard’s 8 English test sets, our INT8 models reach the published accuracy of the full-precision originals. Canary-Qwen 2.5B scored 5.62% word error rate against the published 5.63%, and Whisper large-v3 scored 7.22% against 7.44%.</p><h3>1-Click Installation</h3><h3>What You Download</h3><p>You get one small zip file with the installers for every platform.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*PVT4s46TXnwUlJHh.png" /></figure><p>Extract it into any folder. On Windows, double-click Windows_Install_Update.bat, then start the app with Windows_Start_app.bat. Run the same installer again at any time to update.</p><h3>Latest PyTorch, CUDA 13 and Precompiled Libraries</h3><p>The installer makes its own Python 3.12 virtual environment, so your other apps stay untouched. It installs PyTorch 2.13 with CUDA 13 and our precompiled Flash Attention, xFormers, SageAttention and Triton for Windows. You never compile anything.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*E97AvPivzskECl6e.png" /></figure><p>Every package version is tested with the app. This is the end of a real fresh install on our PC:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*e9MqtH0Olart2yVF.png" /></figure><p>At the end, the installer downloads the speaker label models from our mirror. You do not need a Hugging Face token or any model approval.</p><h3>Start the App</h3><p>Double-click Windows_Start_app.bat. CMD shows every startup step with its time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mK5xknjgQAQgXZYu.png" /></figure><p>On our RTX 5090 the app was ready in 6.5 seconds. Open the local address in your browser and start transcribing.</p><h3>Cloud GPUs: RunPod, SimplePod and Massed Compute</h3><p>You can also run the app on a cloud GPU. The cloud installers set up Python 3.12, FFmpeg n9.0 and everything else with one command, and a Gradio share link lets you use the app from any device.</p><ul><li>SimplePod: <a href="https://simplepod.ai/ref?user=secourses">register here</a> and use <a href="https://dash.simplepod.ai/account/explore/100/ref-secourses/">this template</a>.</li><li>RunPod: <a href="https://get.runpod.io/955rkuppqv4h">register here</a> and use <a href="https://get.runpod.io/SECourses_CU13">this template</a>.</li><li>Massed Compute: <a href="https://vm.massedcompute.com/signup?linkId=lp_034338&amp;sourceId=secourses&amp;tenantId=massed-compute">register here</a> and use our coupon SECourses.</li></ul><p>The step-by-step commands are in the instruction files inside the zip.</p><h3>Requirements</h3><p>On Windows you need Python 3.12, Git, FFmpeg, CUDA 13, cuDNN 9.17 and Visual Studio with C++ tools. <a href="https://youtu.be/DrhUHnYfwC0">This tutorial</a> shows every step, and the same setup runs all our AI apps. The app runs on NVIDIA GPUs from the GTX 16 and RTX 20 series up to the RTX 50 series. Our INT8 ConvRot engines use the RTX 30 series and newer, and other GPUs switch to the standard models automatically.</p><h3>Config Presets</h3><p>Every engine comes with a locked best-quality preset. We researched each value on real test sets, so you get top results without changing a single setting.</p><p>Where to find it: Config Presets sits at the top of the page and stays there on every tab.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sS9P5eM0z0xoWwJe.png" /></figure><p>Open the Select Preset list to see every preset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*D8POK5Qo0ntjFqVK.png" /></figure><p>Change anything you like, type a name and click Save to keep your own preset. The app remembers your last preset and loads it at every start.</p><h3>Three Engines and 21 Whisper Models</h3><p>Choose the engine in Base Model: Whisper (faster-whisper), Insanely Fast Whisper (Transformers) or NVIDIA Canary-Qwen 2.5B. Each engine loads its best settings when you select it.</p><p>Where to find it: open the File tab and scroll to Base Model. The Model list is right under it, and the Youtube and Mic tabs have the same controls.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hpTnyPRZIAWUwHMG.png" /></figure><p>Open the Model list to see every model:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Z6CunhV0FY7bQDMG.png" /></figure><p>You get every Whisper model from tiny to large-v3, plus turbo, distil and our two INT8 ConvRot models. A model downloads automatically the first time you use it.</p><h3>NVIDIA Canary-Qwen 2.5B</h3><p>Canary-Qwen 2.5B is one of the most accurate open English speech models on the Open ASR Leaderboard (September 2026). Our INT8 ConvRot build runs it with Triton kernels and CUDA graphs.</p><p>Where to find it: pick Canary Qwen Best Quality in Select Preset, or choose Canary-Qwen (NVIDIA NeMo) in Base Model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*APKUJWJDL2kqB4Vg.png" /></figure><p>The Canary settings then load automatically:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*P10Ub6APWK67Zdoy.png" /></figure><p>The app reads your GPU memory at startup and picks the batch size for you: batch 2 on 6 GB, 4 on 8 GB, 8 on 10 to 12 GB and 16 on 16 GB and more. Here it chose batch 16 on the 32 GB RTX 5090 and finished the 1 hour 29 minute lecture in 36 seconds.</p><h3>Automatic Model Downloads</h3><p>You never download models by hand. The first time you use a model, the app downloads it and shows the progress in Live Transcription and in CMD.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*S1cypMky3JSvwWud.png" /></figure><p>All models and caches stay inside the app’s own models folder, so the rest of your PC stays clean.</p><h3>100 Languages and Translation to English</h3><p>Whisper understands 100 languages, and Automatic Detection finds the language for you.</p><p>Where to find it: in the File tab, the Language list and the Translate to English checkbox sit next to the Model list.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*H69QahoEvQEaBorr.png" /></figure><p>Open the Language list to see every language:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qyWcQlbGS7QtOo0v.png" /></figure><p>Turn on Translate to English and Whisper writes English subtitles directly from speech in any language.</p><h3>Output Formats and Run Controls</h3><p>All the main controls sit in one clear panel under the model settings.</p><p>Where to find it: in the File tab, right under the Model row. Tick your formats, then click GENERATE SUBTITLE FILE.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SPH2Gxh99uCDQ2Zw.png" /></figure><p>Here are the formats and run controls up close:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jWZaLfJCKD2cvOf5.png" /></figure><p>1: tick as many formats as you like. 2 and 3: start and stop jobs. 4: previous-text context is tuned for long files automatically. 5: open the extra sections for advanced settings, the music remover, the voice filter and speaker labels. 6: every job ends with a clear summary. One run writes all six files, named after your input file:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*MUALRGPzKKEUv5Vs.png" /></figure><p>The best-quality presets use word timestamps. You get clean sentence-level subtitles, or word-level highlighted SRT and WebVTT when you want them.</p><h3>Advanced Parameters</h3><p>Where to find it: open the File tab, scroll down and click Advanced Parameters. The Open / close all sections button at the top opens it too, together with every other section.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NyFilY48c2tD_vsR.png" /></figure><p>The panel opens with every decoding setting, each with a short explanation under it:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ecKYumzwoEobTNcO.png" /></figure><p>Use Hotwords or the Initial Prompt for names and terms. Turn on Use Batched Inference for extra speed on long files. Offload Models to RAM When Idle keeps your models ready while your VRAM stays free.</p><h3>Music Remover, Voice Filter and Speaker Labels</h3><p>Three built-in filters prepare your audio before transcription.</p><p>Where to find it: in the File tab, scroll down. The three sections sit right under Advanced Parameters; click a title to open it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*luwN7-cv0WSEL5ab.png" /></figure><p>Here are the three sections opened:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LDpzlB0SfS4ddPtu.png" /></figure><p>The Background Music Remover separates speech from music with UVR MDX-Net. The Silero voice filter skips silence. Diarization adds speaker labels such as SPEAKER_00 and SPEAKER_01 without a Hugging Face token. In our test it labelled every turn of a two-person interview correctly.</p><h3>Batch Processing Whole Folders</h3><p>Transcribe a whole folder in one click, subfolders included.</p><p>Where to find it: in the File tab, on the right side, next to the upload box.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Zsu7xsghqjm_JzU0.png" /></figure><p>Tick Enable Batch Processing and set your folders:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-ixpB0YMAr-j-cLO.png" /></figure><p>Here 3 videos in 3 subfolders, 18 minutes in total, were finished in 26 seconds. The output folder mirrors your folder tree, finished files are skipped on the next run, and one broken file never stops the batch.</p><h3>YouTube Videos and Whole Channels</h3><p>Paste a YouTube link and the app loads the thumbnail, title and description.</p><p>Where to find it: click the Youtube tab. The Youtube Link box is at the top.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*biD0sProL79IJyL9.png" /></figure><p>The link loads the video details right away:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*en6f82owadM2T7od.png" /></figure><p>Click Generate and the app downloads the audio and transcribes it. Our 12 minute video went from link to finished subtitles in about 15 seconds. Turn on Mass Transcribe Latest Channel Videos to process the latest videos of a whole channel, one after another.</p><h3>Live Microphone</h3><p>Speak and watch the text appear. The Mic tab has two modes.</p><p>Where to find it: click the Mic tab. Live Mic is on the left and Record Then Generate on the right.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*c2GREgPkTg9a2RCW.png" /></figure><p>Press Live Mic Record and start speaking:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9zxFvdy4jTS79kGx.png" /></figure><p>Live Mic shows a preview while you speak and updates it every 2 seconds. When you stop, the whole recording is transcribed with the full model and saved as subtitle files. Record Then Generate lets you record first and create the subtitles after.</p><h3>Subtitle Translation to 200 Languages</h3><p>Translate your subtitle files with Meta NLLB, right inside the app.</p><p>Where to find it: click the T2T Translation tab, drop your subtitle files, then open the NLLB tab. DeepL API is right next to it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iLQ2QoAFyIkLUvF_.png" /></figure><p>Choose the languages and click TRANSLATE SUBTITLE FILE:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*XlQYPIuDPjAqHqDg.png" /></figure><p>Choose the source and target language, click Translate and download the new file. Timestamps and speaker labels stay in place. A DeepL API tab is included too, for your own DeepL key.</p><h3>Background Music Separation</h3><p>The BGM Separation tab splits any audio file into music and voice.</p><p>Where to find it: click the BGM Separation tab, drop your audio files and click SEPARATE BACKGROUND MUSIC.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*X83_h0H__yC2BTQ7.png" /></figure><p>After the separation you get two tracks:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5k9D4CWWgLm3FVK9.png" /></figure><p>You get a music-only track and a clean voice track, both saved in your outputs folder.</p><h3>Light and Dark Theme</h3><p>The app starts in dark mode, and one click switches to the light theme.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qW33C6nB87XyH6mM.png" /></figure><p>The Open / close all sections button opens or closes every panel at once.</p><h3>Latest Updates</h3><p>Whisper-WebUI Premium gets frequent updates. Versions 12.4 to 13 came out between 27 and 29 September 2026:</p><ul><li>New INT8 ConvRot models for Whisper large-v3, large-v1 and Canary-Qwen 2.5B, downloaded automatically.</li><li>Lower word error rate in English: the INT8 models now match the published accuracy of the originals.</li><li>Canary-Qwen transcribes recordings up to 40 seconds in one piece and cuts longer ones at pauses.</li><li>Canary batch size is picked from your GPU memory, with automatic recovery when VRAM runs short.</li><li>Models wait in RAM between jobs, so the next job starts in under a second.</li><li>Faster start, with every startup step shown in CMD.</li><li>Batch processing keeps your subfolders, skips finished files and lists any failed files at the end.</li><li>All GPU jobs wait in one queue and run one after another, so two jobs never clash.</li><li>A redesigned interface with light and dark themes.</li></ul><h3>Get Whisper-WebUI Premium</h3><p>Download the latest zip file attached to this post, extract it and run the installer for your platform. To update, get the newest zip, overwrite the old files and run Windows_Install_Update.bat again.</p><ul><li>Full tutorial video: <a href="https://youtu.be/yzNhkC3aQuA">watch it on YouTube</a>.</li><li>Requirements tutorial: <a href="https://youtu.be/DrhUHnYfwC0">Python, Git, FFmpeg, CUDA and C++ tools step by step</a>.</li><li>Support: <a href="https://discord.com/channels/772774097734074388/1079506787734134844">our Discord channel</a>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YUxSYRXQY0PadCF0e5e5Lg.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0d1740d6b59a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>