<h1 id="heading-426">ruru.XL</h1><h2 id="heading-427">Hello! Introducing LoRA, the SDXL model from the ruru series.</h2><p>LoRA is an anime-oriented model within the SDXL family, capable of producing even more delicate, adorable, and beautiful outputs. <br />It's designed with higher steps and resolutions in mind, so please be prepared for longer processing times. <br />But I'm sure you'll notice a distinctive difference compared to what we've had before in an instant!</p><h2 id="heading-82">Base Model:</h2><ul><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Linaqruf/animagine-xl"><strong>Animagine(v1.0)</strong></a></p></li></ul><h3 id="heading-428">Recommended Settings</h3><ul><li><p>Step: 55 or higher</p></li><li><p>Resolution: 1024 x 1024 or higher</p></li><li><p>Sampler: Euler a / DPM++ 2S a Karras (Of course, other options are welcome too! We look forward to hearing your feedback!)</p></li><li><p>TI: <a target="_blank" rel="ugc" href="https://civarchive.com/models/118418/negativexl"><strong>negativeXL</strong></a></p></li><li><p>Be cautious with <strong>{{{{nsfw}}}}</strong> content</p></li><li><p>Reducing negativity and increasing positivity may introduce a bit more randomness into the outputs.</p></li></ul><h1 id="heading-429"><span style="color:rgb(209, 213, 219)">Enjoy this model!</span></h1>

ruru.XL

<p>This time, we made some significant changes. One of the notable differences is the number of steps. Previously, we used my home's RTX 4090, but this time, I borrowed an H100 from a friend.<br />The base remains Animagine-XL.</p><ul><li><p>Updated the dataset (from WD14-moat to ML-Danbooru).</p></li><li><p>Optimized the settings (approximately three times the number of steps for the same epoch).</p></li><li><p>Optimal strength: 0.6 to 0.8.</p></li><li><p>DPM++ 2Sa Karras is the best choice.</p></li><li><p>Negative-XL is not necessary (please refer to Animagine-XL for guidance).</p></li></ul>

Files with SHA256: dbc7f9c28d1dc7a33373bb939ef6dabe8c1347d6d89a9b560822090d7ccecdd4