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: training config, training dataset config, "eval" dataset config (stabilized training loss).
Captioning script: link.
Dataset construction process:
Select and download images manually from the internet.
Use Flux 2 Klein 9b ComfyUI workflow to batch process the whole folder and remove watermarks.
Caption with Gemma 4 31b. If you have less VRAM, JoyCaption or one of the medium sized Qwens would work almost as well.
Main training parameters:
Rank 32 LoRA
Global batch size 4
Don't train LLM adapter (llm_adapter_lr=0)
AdamW optimizer, 2e-5 LR, betas=[0.9, 0.99]
Mixed res training on [512, 1024, 1536]
[512, 1024] mixed is always good, 1536 is optional but helps acquire details for this particular style.
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.
This version corresponds to 40 epochs (120 passes over the data when considering the 3 resolutions)