My Character <a target="_blank" rel="ugc" href="https://huggingface.co/spaces/waveshaker/loras_gallery">LoRA Gallery</a> (1000+ Downloadable Character LoRAs &amp; 500+ Runnable Online)<hr />Anima:<ul><li>Steps: 16450</li><li>Effective BS: 4 (4 batch × 1 grad accum)</li><li>LR: Scaled up for BS 4 (≈ 1e-04 @ BS 1)</li><li>Dataset: Same as Krea 2 (see Krea 2 page for download)</li></ul>Moss Image:<ul><li>Steps: 5460</li><li>Effective BS: 10 (10 batch × 1 grad accum)</li><li>LR: Scaled up for BS 10 (≈ 1e-04 @ BS 1)</li><li>Dataset: Same as Krea 2 (see Krea 2 page for download)</li></ul><hr />Collected nearly all works of Christie's on DeviantArt, and processed the images through a comprehensive combination of Qwen+Krea2+Klein. The style of the final dataset images falls between 3D CG and real photography. The final dataset consists of 3159 images, and it's available for download —— <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/3350472?fileId=3237863">traindata(1.28G)</a>The model has plenty of knowledge regarding NSFW, poses, and streethookers, but it is also prone to causing artifacts. The dataset ratio for training faces versus my niche/guilty pleasures was roughly 7:3. Feel free to explore the dataset's potential and try out different concepts.<h2 id="krea2-or-moss-image">KREA2 | Moss Image</h2>I personally recommend using lodestones' <a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/Kroma/tree/main">Kroma v0.3</a>. Due to its training dataset, it does a great job of restoring Christie's features. I also recommend using <a target="_blank" rel="ugc" href="https://civarchive.com/models/2727636/realistic-slider-krea-2-rank1-7mb?modelVersionId=3066073">Realistic Slider - (Krea-2) (Rank1 7mb)</a>, which allows you to easily switch between photographic realism and 3DCG styles. Testing shows that this LoRA performs better on Kroma than on other fine-tuned models.

Files with SHA256: 2c78ee0cb92647097dba4c5bc8eae4c906fa044d9c6cd2e344bd91228936f4bf