Greg Rutkowski style LoRA for Anima. Trained on preview3. Prefix prompt with "@greg rutkowski. " Natural language prompts work best.All training data and diffusion-pipe training config files are shared in the attached ZIP file. If you just want the config files: <a target="_blank" rel="ugc" href="https://gist.github.com/tdrussell/3f79596efb8e27672da0881afd9c1d51">training config</a>, <a target="_blank" rel="ugc" href="https://gist.github.com/tdrussell/d5651bcf4a3f0855c7f55f5148519d73">training dataset config</a>, <a target="_blank" rel="ugc" href="https://gist.github.com/tdrussell/6a1eea3aa0634d3d02c37462fefe5b9f">"eval" dataset config</a> (stabilized training loss).Captioning script: <a rel="ugc" href="https://gist.github.com/tdrussell/37f3099a34f1ee7343c7ae321feb1aa5">link</a>.Dataset construction process:<ol><li>Select and download images manually from the internet.</li><li>Use Flux 2 Klein 9b ComfyUI workflow to batch process the whole folder and remove watermarks.</li><li>Caption with Gemma 4 31b. If you have less VRAM, JoyCaption or one of the medium sized Qwens would work almost as well.</li></ol>Main training parameters:<ul><li>Rank 32 LoRA</li><li>Global batch size 4</li><li>Don't train LLM adapter (llm_adapter_lr=0)</li><li>AdamW optimizer, 2e-5 LR, betas=[0.9, 0.99]</li><li>Mixed res training on [512, 1024, 1536]<ul><li>[512, 1024] mixed is always good, 1536 is optional but helps acquire details for this particular style.</li></ul></li><li>sigmoid_scale=1.3, optional but helps with fine details by shifting the timestep distribution towards something that is slightly more uniform than standard logit normal.</li><li>This version corresponds to 40 epochs (120 passes over the data when considering the 3 resolutions)</li></ul>

Files with SHA256: dad71dff45741e1d5ca9bffaf30e608aa41d3b9cc2b641bf4b145de4ad85b28d