<?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 Wiro AI on Medium]]></title>
        <description><![CDATA[Stories by Wiro AI on Medium]]></description>
        <link>https://medium.com/@wiroai?source=rss-d3442d2c5112------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Q1nzgDSb_l_IRdY6Ig0rAw.png</url>
            <title>Stories by Wiro AI on Medium</title>
            <link>https://medium.com/@wiroai?source=rss-d3442d2c5112------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 11:19:28 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@wiroai/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[We made our launch film without a camera. Five things broke.]]></title>
            <link>https://medium.com/@wiroai/what-broke-when-we-made-a-106-second-founder-film-with-ai-models-7771d0411e96?source=rss-d3442d2c5112------2</link>
            <guid isPermaLink="false">https://medium.com/p/7771d0411e96</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[filmmaking]]></category>
            <category><![CDATA[video-production]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Wiro AI]]></dc:creator>
            <pubDate>Sat, 21 Nov 2026 19:45:33 GMT</pubDate>
            <atom:updated>2026-11-21T20:58:13.146Z</atom:updated>
            <content:encoded><![CDATA[<h4>Skin, voice, music, timing and an AI judge: what went wrong in our 1:46 founder film, and what fixed it</h4><p><em>This story was written with the assistance of an AI writing program, from our team&#39;s production notes.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/803/0*3uPoTKOGKlEgpKJD.jpg" /><figcaption>A frame from the film. Every shot in it was generated with AI models.</figcaption></figure><p>Here is the film: <a href="https://www.youtube.com/watch?v=jx1MJtBjIyg">watch it on YouTube (1:46)</a>.</p><p>In our Product Hunt launch film, our founder walks down a street, works late and speaks straight to the camera. None of it was filmed. Every shot is generated, the voice is a clone, and the screen animations are an HTML page rendered one frame at a time.</p><p>We finished it last week, after five versions and about 800 model runs. Most of the work went into problems that looked fine in a single clip and fell apart once the clips were cut together. Here are five of them, and what we did about each.</p><h3>Sharpening made him look older</h3><p>The raw video came out soft, so we ran every shot through an upscaler. It sharpened the skin too: lines, stubble shadow, a darker tone. The founder’s note on version 4 was short: hands and skin look old, wrinkled and too dark.</p><p>Our answer was to build each frame from two sources. Skin-colored areas come from the raw frame, and the rest comes from the sharpened one. The skin is then lifted toward the measured brightness of the face, capped at +24%, with the eyes and brows protected. He chose the strength himself from a sheet of four columns placed side by side. One shot’s warm brick wall came out a bit softer as well. He looked at it and was fine with it.</p><h3>The more we fixed the voice, the worse it got</h3><p>The cloned voice sounded good in version 1, used exactly as it came out. In version 2 we time-stretched lines so they would fit the lips. In version 3 we cleaned the voice sample and ran EQ and a de-esser. Each step made it worse. First it sounded muffled, then echoey, then like a phone call.</p><p>The answer was to leave it alone. We kept a single change, made to the clone sample itself: the source recording had nothing above 11 kHz, so we restored that missing top band before cloning. Whatever comes out of generation now stays raw.</p><h3>Five songs where we wanted one</h3><p>The first soundtrack was built in chunks from a composition plan. It came out sounding like five different songs. The music in the film is a single track from a single prompt, and we don’t stretch it. We moved the film’s pauses onto the track’s bar lines instead, and where it ran long we cut out whole 4-bar loops.</p><h3>Moving the picture to match the voice</h3><p>The talking shots never matched the timing of the narration. Stretching audio ruins it, as the voice had already shown us, so we did the opposite and changed the video. Forced alignment gives us word and character timings for the clip’s own audio and for our narration line, and the video is warped through those anchor points.</p><p>One thing surprised us. In the generated talking clips, the lips ran about 90 ms ahead of their own audio. That figure comes from 34 mouth closures measured across 11 clips. We shift the picture by about 60 ms, which leaves the lips just slightly early, and that reads as natural.</p><p>We also generated 10 or more takes of every talking shot and kept the one whose pace matched the line. A standing shot can hide a shortened pause. A walking shot shows it.</p><h3>An AI judge with blind spots</h3><p>To compare candidates we used a multimodal model. With still images it was reliable. In audio A/B tests it favored whichever clip came first, so we only trusted a result that held up in both orders. It also judged a 400 ms lip-sync offset to be “in sync”. So lip sync was checked by eye, using frame strips at every m, p and b.</p><h3>What we’re taking into the next film</h3><ul><li>Every part of the project kept: clips, timings, music and the render page. One scene can change without redoing the film.</li><li>A playbook started on day one. Each of the founder’s notes became a rule, so the next film begins with those rules instead of from zero.</li><li>Looking at the frames themselves, not at a summary of them.</li></ul><p>We ran every model through <a href="https://wiro.ai/">Wiro</a>, our own API for image, video, audio and language models. That made it easy to switch from one model to another halfway through.</p><p><strong>P.S.</strong> On December 3 we launch Wiro on Product Hunt. If you enjoyed this story, stop by, say hi and tell us what you think: <a href="https://www.producthunt.com/products/wiro-ai">Wiro AI on Product Hunt</a>.</p><p><em>Originally published at </em><a href="https://dev.to/wiro/what-broke-when-we-made-a-106-second-founder-film-with-ai-models-50nn"><em>https://dev.to</em></a><em> on November 21, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7771d0411e96" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>