<?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 Shahryar Saqib on Medium]]></title>
        <description><![CDATA[Stories by Shahryar Saqib on Medium]]></description>
        <link>https://medium.com/@sabreking?source=rss-fe61294ab31b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*hr_UC2XH6jOHXgVGWJrsbQ.png</url>
            <title>Stories by Shahryar Saqib on Medium</title>
            <link>https://medium.com/@sabreking?source=rss-fe61294ab31b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 18:32:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sabreking/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[The New AI Powered Art Pipeline for Games-Part 1]]></title>
            <link>https://medium.com/@sabreking/the-new-ai-powered-art-pipeline-for-games-part-1-5d1898d930c4?source=rss-fe61294ab31b------2</link>
            <guid isPermaLink="false">https://medium.com/p/5d1898d930c4</guid>
            <category><![CDATA[game-production]]></category>
            <category><![CDATA[game-development]]></category>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[art-department]]></category>
            <category><![CDATA[generative-ai]]></category>
            <dc:creator><![CDATA[Shahryar Saqib]]></dc:creator>
            <pubDate>Tue, 06 Jun 2023 11:26:23 GMT</pubDate>
            <atom:updated>2023-06-06T11:26:23.445Z</atom:updated>
            <content:encoded><![CDATA[<h3>Game-Changer: The New AI-Driven Art Pipeline for Games — Part 1</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KVnw4FFj28wwobvIPx51oA.png" /></figure><p>Due to the advent of AI generative models and the enthusiasm with which the open-source community has adopted and run with them, I’ve made some realizations about their implications for the game development market. This led to the development of a new workflow for game development.</p><h3>Realizations</h3><p>Training custom art generation models has become incredibly easy. It has become cheap and simple to train a model that will output art and assets which can be exactly what the client’s or game’s art style needs. You can do it on one high-powered laptop. With good prompting, the entire art requirements of a game can be met with one AI model trained specifically for that game. The new art departments of game development companies will consist of prompt engineers, art directors, and sketch artists (for Control nets — explained later).</p><p>This journey started as an experiment on my personal laptop to validate these assumptions, and I was blown away by how successful it turned out to be. At this stage, I am confident enough to say that I can train a model and churn out art for a whole medium scope game in just one week.</p><p>The pipeline I’ve developed follows several steps. I will cover each of these in detail, step by step, across multiple parts of this series of posts.</p><p>the steps will cover high level concepts and “what to do”s it wont go into details on “how to do them exactly” since that part can basically be covered by a motivated engineer/artist in a couple of days following youtube tutorials etc in order to fill the role of a “Prompt Engineer”.</p><h3>Step 1: Understanding the Art You Need</h3><p>Gather references in collaboration with the art director or the client. Use CLIP/BLIP (AI interrogation of those images to isolate what that art style actually is in terms that AI will understand) on the references.</p><p>Interrogation of the reference images using a CLIP model basically tells you how the AI will describe the image, giving you insight into what your references can be described as.</p><p>Examples of the kind of data you will get:</p><pre>&gt;a purple and black picture of a tall tower in the middle of a forest with a purple light coming from it&#39;s top and a purple light shining on the top, (Filip Hodas:0.724), (league of legends splash art:0.104), (concept art:0.246), (fantasy art:0.169)<br>&gt;a doorway in a dark forest with a glowing light coming from the entrance to the entrance to the room with plants growing out of it, (Filip Hodas:0.427), (rossdraws global illumination:0.337), (a 3D render:0.361), (fantasy art:0.241)<br>&gt;a painting of a river flowing through a lush green valley at night with lights on the mountains and a river running through the valley below, (Chris LaBrooy:0.289), (cinematic matte painting:0.276), (a detailed matte painting:0.629), (fantasy art:0.649)</pre><p>Copy and paste all the outputs of CLIP interrogation into a text file. Use ChatGPT4 to crunch that data in terms of what terms, tags, words, colors, lighting descriptions, artists, etc are being used in those reference images and how much and convert it into multiple CSV files:</p><p>Here is an example of the data i crunched from interrogation of about 14 reference images.</p><pre>Tags occurrence<br>Tag,Average Weight,Occurrence Percentage<br>league of legends splash art,0.104,4.00<br>concept art,0.228,12.00<br>fantasy art,0.256,32.00<br>rossdraws global illumination,0.286,12.00<br>3D render,0.386,24.00<br>cinema 4 d,0.227,8.00<br>cubo-futurism,0.418,4.00<br>matte fantasy painting,0.082,4.00<br>space art,0.441,8.00<br>volumetric lights,0.083,4.00<br>egyptian art,0.660,4.00<br>holography,0.231,8.00<br>retrofuturism,0.341,12.00<br>rendered in octane,0.211,4.00<br>environmental art,0.290,4.00<br>unreal engine render,0.076,4.00<br>intricate environment,0.176,4.00<br>bioluminescence,0.051,4.00<br>fireflies,0.129,4.00<br>tilt shift photo,0.477,4.00<br>detailed matte painting,0.447,12.00<br>matte painting,0.456,8.00<br>cinematic matte painting,0.272,8.00<br>Artist occurrence:<br>Artist,Average Weight,Occurrence Percentage<br>Filip Hodas,0.585,40.00<br>Chris LaBrooy,0.581,12.00<br>Anton Fadeev,0.433,12.00<br>Anato Finnstark,0.580,4.00<br>Beeple,0.303,8.00<br><br>Keywords occurrence in descriptions of this world<br>Keyword,occurrencePercentage<br>tower,2.02<br>forest,7.07<br>light,8.08<br>purple,3.03<br>doorway,2.02<br>dark,2.02<br>glowing,2.02<br>plants,2.02<br>river,3.03<br>valley,2.02<br>mountains,2.02<br>Waterfall,2.02<br>night,3.03<br>lights,2.02<br>computer,2.02<br>city,2.02<br>landscape,2.02<br>building,2.02<br>lake,2.02<br>rocks,3.03<br>Trees,4.04<br>painting,4.04<br>Colour use percentage.<br>Colour,Percentage<br>Purple,24%<br>Green,20%<br>Blue,16%<br>Orange,8%<br>Yellow,12%<br>Black,12%<br>White,8%</pre><p>Get ChatGpt to decribe the salient features of the imaginary world</p><pre>Nature-rich: Diverse landscapes with forests, mountains, and water bodies.<br>Fantastical: Supernatural elements, glowing lights, and magical objects.<br>Futuristic: Blend of technology and nature, cybernetic elements, and futuristic architecture.<br>Visually striking: Emphasis on light sources, vivid colours, and atmosphere.<br>Surreal: Dreamlike quality with unusual structures and combinations.<br>Diverse architecture: Mix of traditional, ancient, and futuristic styles.</pre><p>Use ChatGpt (preferably 4) to generate more prompts. An example of a prompt you can use (this is in no way perfect but it got me the proof of concept):</p><pre>generate a list of 10 prompts to generate a beautiful image from this world, the format of the prompt should be like this.&lt;prompt&gt;,&lt;artist&gt;,&lt;tag(avg-weight)&gt;,&lt;tag(avg-weight)&gt;,&lt;tag(avg-weight)&gt;<br>Remember to take guidance form the occurrence of keywords to generate your prompts ,those keywords will tell you what are the images usually about.Take guidance from the CSVs of artists and tags to decide which artist and tags to use,the occurrence rate of artists and tags should be your guide in choosing which artist and tag to use.Take guidance from the colour use csv to add some colour information to the prompts as well.</pre><p>You will get a response something like this:</p><pre>A glowing waterfall flowing through a lush forest, surrounded by mysterious floating islands, (Filip Hodas:0.670), (fantasy art:0.230), (nature-rich:0.450), (surreal:0.330)<br>An ancient temple illuminated by neon lights, nestled within a futuristic cityscape, (Chris LaBrooy:0.585), (retrofuturism:0.365), (diverse architecture:0.420), (visually striking:0.380)<br>A cybernetic tree in a dreamlike forest, adorned with magical bioluminescent flowers, (Anato Finnstark:0.580), (fantastical:0.280), (futuristic:0.250), (nature-rich:0.450)<br>A peaceful lake reflecting a vibrant sunset, with technologically advanced structures emerging from the water, (Anton Fadeev:0.484), (futuristic:0.310), (visually striking:0.390), (nature-rich:0.450)</pre><p>Use the prompts to generate even more images that are similar to what you need. Set the CFG scale to low (2–5); we want variation at this stage. Set the size to 512x512 to make the next step easier. Gather all of the images that you generated and build a dataset out of them. Get the art director or client to shortlist the images that you generated and remove those that they don’t ‘like’ or feel aren’t the art style they want. In the end, we need about 50-250 images for our new training set.</p><p>There’s no need to BLIP caption the images for the training dataset, as we’re going to be training a STYLE as opposed to a SUBJECT, therefore we do not want a strong correlation with a specific tokens except the keyword/trigger.</p><p>Train a LORA or a whole checkpoint to generate art for use in your game.</p><h3>Caveats</h3><p>Most of these instructions assume that you or have an engineer on your team who has a basic understanding of using Stable Diffusion WebUI and that you have a good prompt engineer on your team.<br>Do this whole process separately for UI elements (icons, buttons, etc.) to get a separateUI model and train concepts like button, icon in that model.</p><p>The following sample images are far from perfect but show the potential of what can be achieved to meet the need for your projects</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KVnw4FFj28wwobvIPx51oA.png" /><figcaption>Some Game Icons Generated through this process</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Hxlr6rr3ygiwu9_1hnYUwA.png" /><figcaption>Characters and Icons</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QZ1OrMQTrDtj2T-zrCl9cg.png" /><figcaption>Seamless Texture Generation for Games</figcaption></figure><h3>Coming Up Next in the Series</h3><p>Stay tuned for Part 2 of this series where we will further explore our innovative game art pipeline. We’ll delve into techniques like using control nets and sketch artists to generate art that perfectly aligns with your vision. We’ll also cover how to modify existing art for expanded uses — think of creating an injured version of your warrior sprite or a broken version of your wall texture.</p><p>Feel free to engage with this article and comment if you are interested in Part 2 and if there is something specific you want me to cover.</p><p>In Part 3, our main focus will be on how you can hire and develop job descriptions for the future art departments in the gaming industry, particularly emphasizing roles that collaborate with AI tools. Until then, happy coding and game developing!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5d1898d930c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>