<h1 id="heading-11756">Kohaku XL Gamma</h1><p>A SDXL anime base model aims to create unique artworks.</p><p><strong>Not designed for reproducing copyright characters or artist style</strong></p><p></p><hr /><h2 id="heading-11757">Introduction</h2><p>This model can be seen as a derivative of Animagine XL 3.0 project.<br />Basically I'm collaborating with Linaqruf for making better Anime base model (and it is obvious that we have different goal/target)<br />We share our models and technique to improve our models' quality.<br />And that is also how this model been created.</p><p></p><hr /><h2 id="heading-6931">Models</h2><h3 id="heading-11758">Base7</h3><p>Kohaku-XL base7 is resumed from beta7 and use same dataset that beta series have used. But this time I use my own metadata system to create captions. (Can be taken as advanced version of what linaqruf used, will open source it soon)<br /><br />The metadata database can be downloaded here:<br /><a target="_blank" rel="ugc" href="https://huggingface.co/datasets/KBlueLeaf/danbooru2023-sqlite">KBlueLeaf/danbooru2023-sqlite · Datasets at Hugging Face</a><br /><br />Trainin details:<br />LR: 8e-6/2e-6<br />Scheduler: constant with warmup<br />Batch size: 128 (batch size 4 * grad acc 16 * gpu count 2)</p><h3 id="heading-6932">Base 8</h3><p>Kohaku-XL base8 is as same as base7 but use 5e-6/1e-6 LR, and it have a full epoch on 1.48M images.</p><p></p><h3 id="heading-11759">Gamma rev1</h3><p>Kohaku-XL Gamma rev1 is a merged model which combine the learned diff from anxl3 and kohaku xl base 7. With this forumla:</p><pre><code>gamma rev1 = beta7 + 0.8 * (anxl3 - anxl2) + 0.5 * (base7 - beta7)</code></pre><h3 id="heading-6933">Gamma rev2</h3><p>Kohaku-XL gamma rev2 is merged with this equation:</p><pre><code>gamma rev2 = beta7 + 1.0 * (anxl3 - anxl2) + 0.25 * (base8 - beta7)</code></pre><p>I also use a MBW recipe to merge with Anxl3 directly:</p><pre><code>0,0.1,0.1,0,0.1,0.1,0,0.1,0.1,0,0,0,0,0,0.05,0.05,0.05,0.05,0.05,0.05</code></pre><h2 id="heading-6934"></h2><h3 id="heading-6935">Why merge:</h3><p>First of all, (anxl3 - anxl2) can be take as "What Anxl3 have learned from the training resumed from anxl2". And since the "learned things" are basically "meta tags" and some character/style information, it is good to merge the diff from anxl3/anxl2. </p><p>Or you can take this merge as "merged with 2 lora/lycoris model, one is extracted from anxl3 and another is extracted from base7/8"</p><p>The very low weight MBW merge is to fix some over-trained artifact in the merged model.</p><p></p><hr /><h2 id="heading-11760">Usage</h2><p></p><h3 id="heading-6936">FP8 model</h3><p>After safetensors 0.4.2 or pytorch 2.2.0, users can use fp8 to save their model into safetensors or pytorch ckpt. And I also upload a fp8 version of the model in zip, you can unzip it and drag the fp8 safetensors into your sd model folder and use it normally.</p><p>Remember to update your safetensors to 0.4.2.</p><p></p><p>More information of FP8+FP16 inference:</p><p><a target="_blank" rel="ugc" href="https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14031">A big improvement for dtype casting system with fp8 storage type and manual cast by KohakuBlueleaf · Pull Request #14031 · AUTOMATIC1111/stable-diffusion-webui (</a><a target="_blank" rel="ugc" href="http://github.com">github.com</a><a target="_blank" rel="ugc" href="https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14031">)</a></p><p></p><h3 id="heading-6937">Parameter</h3><p>This model is trained under 768x1024 to 1024x1024 ARB. It is recommended to use pixel count within 786432 ~ 1310720.<br />Recommended CFG scale is 4~7.<br />Sampler should not matter.</p><p></p><h3 id="heading-6938">Tagging</h3><p>This model use my own system for quality tags or something like that. <br />So although this model combine the diff weight from anxl3, I will still recommend user to use mine (or both) tagging system.<br /></p><p>This model is trained with artists' name as a tag so you can use artists tag as "stylize tag". But since my training config is not designed for learning style. I don't think you can actually reproduce the style of any specific artists. Just use artists tag as style refiner tag.<br />For character tags, as same as artists. The model can accept character tags but I won't ensure its ability on reproducing any character.</p><p><br />The format of prompt is as same as anxl3. (You can check the sample images I post)<br /><br />Rating tags:</p><ul><li><p>General: <code>safe</code></p></li><li><p>Sensitive: <code>sensitive</code></p></li><li><p>Questionable: <code>nsfw</code></p></li><li><p>Explicit: <code>explicit, nsfw</code></p></li></ul><p><br />Quality tags (Better to worse):</p><ul><li><p>Masterpiece</p></li><li><p>best quality</p></li><li><p>great quality</p></li><li><p>good quality</p></li><li><p>normal quality</p></li><li><p>low quality</p></li><li><p>worst quality</p></li></ul><p><br />Year tags (New to Old):</p><ul><li><p>newest</p></li><li><p>recent</p></li><li><p>mid</p></li><li><p>early</p></li><li><p>old</p></li></ul><p></p><p>You may meet some subtle mosaic-like artifact, that may be caused by high-lr or bad resizing/image encoding.<br />I will try to fix it in next version. For now, try to use R-ESRGAN anime6b or SCUNet models for fixing it.</p><p></p><hr /><h2 id="heading-11761">Future plan</h2><p>Since my dataset have some resize/webp artifacts that will harm the models. I will recreate my dataset based on my new system (and opensource it once I done it).<br /><br />The next plan is to train model on larger (3M~6M) dataset with better configuration (which will require A100s and I plan to spend about 2000~10000 USD on it, if you like my works, consider to sponsor me via buy-me-a-coffee or some BTC-sutff, you can get the link from my github profile: <a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf">KohakuBlueleaf (Kohaku-Blueleaf) (</a><a target="_blank" rel="ugc" href="http://github.com">github.com</a><a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf">)</a>)</p><p></p><hr /><h2 id="heading-664">Appendix</h2><p>When I try to generate some sample images, I found my merge method actually goes wrong at first (due to some bugs in LyCORIS' utils).<br />Although the final version of rev1 is not a bugged one, some of sample images I post is actually generated by it.<br />I will upload the bugged one and base7 onto my Huggingface:<br /><a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf">KBlueLeaf (Shih-Ying Yeh) (</a><a target="_blank" rel="ugc" href="http://huggingface.co">huggingface.co</a><a target="_blank" rel="ugc" href="https://huggingface.co/KBlueLeaf">)</a><br /><br />If you want to follow up the progress of next version (or some other projects)<br />Checkout my homepage:<br /><a target="_blank" rel="ugc" href="https://kblueleaf.net/">Kohaku's Homepage (</a><a target="_blank" rel="ugc" href="http://kblueleaf.net">kblueleaf.net</a><a target="_blank" rel="ugc" href="https://kblueleaf.net/">)</a></p><p></p><hr /><h1 id="heading-676">Licence</h1><p><strong>Important</strong></p><p>This model is licensed under faipl-1.0-sd just like anxl3:<br /><a target="_blank" rel="ugc" href="https://freedevproject.org/faipl-1.0-sd/">Freedom of Development (</a><a target="_blank" rel="ugc" href="http://freedevproject.org">freedevproject.org</a><a target="_blank" rel="ugc" href="https://freedevproject.org/faipl-1.0-sd/">)</a></p>

Kohaku-XL gamma

<p>Use base8-lion (full 1 epoch on 1.48M imgs)<br />to do the merge.</p>

Files with SHA256: 38942904e920be7433765920c571303816effa860c334b3fc66068e4b9047e7e