This is the repository for Separate-and-Enhance: Compositional Finetuning for Text2Image Diffusion Models, published at SIGGRAPH 2024.

[Project Page] [Paper]

Build conda environment by running:

conda create -n sepen python=3.10  
conda activate sepen
pip install -r requirement.txt

see src/run_individual.sh for a sample training script.

see src/run_large.sh for a sample training script.

see src/sample.py for refernce.

install clean-fid via pip install clean-fid then refer to src/eval/fid/eval_fid.py for FID evaluation.

We adopt the implementation from A&E. See src/eval/blip/eval_blip.py for BLIP similarity score evaluation.

Clone and build Detic from their official repo. Then move the Python files under src/eval/detic to the cloned folder. See src/eval/detic/eval_detic.py for details.

The 220 concepts we used for the large-scale experiment is at src/concepts/large_scale.py.
The 200 evaluation prompts are at src/concepts/large_test.txt.

Part of our codes is inspired by Custom Diffusion and Attend and Excite.

We leverage Detic and clean-fid for our evaluation.

@inproceedings{bao2024sepen,
    Author = {Bao, Zhipeng and Li, Yijun and Singh, Krishna Kumar and Wang, Yu-Xiong and Hebert, Martial},
    Title = {Separate-and-Enhance: Compositional Finetuning for Text2Image Diffusion Models},
    Booktitle = {SIGGRAPH},
    Year = {2024},
}