CRAFT LoRA for FLUX.2-klein-9B

LoRA weights of CRAFT (SIGGRAPH Asia 2026): Constrained Reward via Attention Fine-Tuning for Subject Personalization without Composed Targets.

[arXiv] | [Project Page] | [Code]

This is the checkpoint used for the results in the paper (rank 64, 3,000 steps).

Usage

git clone https://github.com/Jihun999/CRAFT && cd CRAFT
pip install -r requirements.txt
huggingface-cli download JIHUN999/CRAFT craft_flux2_klein_9b_lora.safetensors --local-dir checkpoints

python inference.py --lora_ckpt checkpoints/craft_flux2_klein_9b_lora.safetensors \
    --refs man.jpg dog.jpg --prompt "A man walking a dog in a park" \
    --labels man dog --categories human animal --output out.png

Defaults: 1024×1024, 4 steps, guidance 1.0. References are raw images (no segmentation needed).

Citation

@article{park2026craft,
  title   = {CRAFT: Constrained Reward via Attention Fine-Tuning for Subject Personalization without Composed Targets},
  author  = {Park, Jihun and Lee, Kyoungmin and Gim, Jongmin and Jo, Hyeonseo and Kim, Jaeyeul and Zou, Han and Zhan, Zhenpeng and Zhang, Yan and Im, Sunghoon},
  journal = {arXiv preprint arXiv:2608.14403},
  year    = {2026}
}

The weights are an adapter for FLUX.2-klein-9B and are subject to its license.