PASDiff: Physics-Aware Semantic Guidance for Joint Real-World Low-Light Face Enhancement and Restoration

Yilin Ni^1,    Wenjie Li^2,    Zhengxue Wang³    Juncheng Li⁴    Guangwei Gao^3,†    Jian Yang³

¹College of Automation, Nanjing University of Posts and Telecommunications, Nanjing, China     ²School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China     ³PCA Lab, School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China     ⁴School of Computer Science and Technology, East China Normal University, Shanghai, China

🚩 Accepted to ECCV 2026

• [ arXiv ] • [ Appendix ] •

⭐ If PASDiff is helpful for you, please consider starring this repo or citing our paper. Thanks! 🤗

• 2026.07.21: Codes, checkpoints, and datasets are released.
• 2026.06.23: This repo is created.

# git clone this repository
git clone https://github.com/YlNi26/PASDiff.git
cd PASDiff

# create new anaconda env
conda create -n pasdiff python=3.8 -y
conda activate pasdiff

# install python dependencies
conda install mpi4py
pip3 install -r requirements.txt
pip install -e .

Download Pre-trained Models:

Download all the necessary pre-trained weights for our PASDiff (except SD2.1) from Google Drive. Put the models in the checkpoints folder.

Since the official Stable Diffusion 2.1 is not available, you could visit ModelScope to download it and use it locally. Put the model in the checkpoints folder as well.

Download WildDark-Face dataset:

Download the real-world low-light face photo set that collected from the Internet called WildDark-Face from Google Drive. If you would like to test on these photos, please put them in the inputs folder.

Quick Inference:

To perform a quick inference, simply run the following command in your terminal:

python inference.py --in_dir ./inputs/test --out_dir ./outputs/result_test

If you want to achieve the exact visual results on the WildDark-Face dataset as shown in our paper, please set the --retinex_type argument to real:

python inference.py --in_dir ./inputs/WildDark-Face --out_dir ./outputs/result_WildDark-Face --retinex_type real

If our work is useful for your research, please consider citing:

@inproceedings{ni2026pasdiff,
    title={PASDiff: Physics-Aware Semantic Guidance for Joint Real-world Low-Light Face Enhancement and Restoration},
    author={Ni, Yilin and Li, Wenjie and Wang, Zhengxue and Li, Juncheng and Gao, Guangwei and Yang, Jian},
    booktitle={ECCV},
    year={2026}
}

If you have any questions, please feel free to reach me out at lewj2408@gmail.com or nixiaolin26@gmail.com.