Krea 2 style fine-tune (experimental)
Full-parameter fine-tune of the Krea 2 Raw DiT (krea/Krea-2-Raw) for a photographic aesthetic shift.
Trained with DiffSynth-Studio, LR 1e-5, 3 epochs, ~3.7k images. Training data included adult content; outputs may be NSFW.
Krea 2 is licensed under the Krea 2 Community License Agreement (https://www.krea.ai/krea-2-licensing). This derivative is distributed under the same terms; commercial use is restricted per that license, and deployers must implement appropriate content-filter measures.
Files
| Path | Format | Use |
|---|---|---|
int8_convrot/step-*_int8_convrot.safetensors | INT8 ConvRot (12.6 GB) | ComfyUI models/diffusion_models/ — same layout as Comfy-Org/Krea-2 int8 convrot files |
checkpoints/step-1650.safetensors | bf16 DiT state dict (24 GB) | DiffSynth-Studio / ComfyUI bf16 |
This is a Raw-class model: use Krea 2 Raw settings (≈52 steps, CFG 3.5–4.5), with the Qwen3-VL-4B text encoder and Qwen-Image VAE. Not compatible with 8-step Turbo settings.