FrDiff

[ICCV'25] Frequency Domain-Based Diffusion Model for Unpaired Image Dehazing

Overview

<img src="./fig/overview.png" width=100%>

Contribution

• We propose a novel frequency domain-based diffusion model for unpaired image dehazing (FrDiff), which is the first study to integrate a DM into unpaired restoration tasks. FrDiff enhances haze removal capabilities by learning amplitude reconstruction from unpaired data, providing inspiration for other unpaired restoration tasks.
• We propose an amplitude residual encoder (ARE) that generates amplitude residuals to bridge the gaps between hazy and clear domains without adding extra parameters, providing supervision for DM training.
• We propose a phase correction module (PCM), designed to eliminate unwanted artifacts by refining the phase spectrum using an attention mechanism.
• Extensive experiments show that FrDiff outperforms existing SOTA methods with fewer parameters and FLOPs

Tranin and Test

1. Clone this github repo

git clone https://github.com/ChengxuLiu/FrDiff.git
cd FrDiff

2. Install dependencies

pip install -r requirements.txt

3. Prepare training dataset and modify "dataroot" and "datafile" in ./train_S1.sh and ./train_S2.sh
4. Run training of stage1

# stage one
sh train_S1.sh

5. The models of stage1 are saved in ./checkpoints and fed into stage2 (modify "pretrained_nameS1" in ./train_S2.sh)
6. Run training of stage2

# stage two
sh train_S2.sh

7. The models of stage2 are also saved in ./checkpoints
8. Prepare testing dataset and modify "dataroot" and "datafile" in ./test.sh
9. Run test

sh test.sh

Citation

If you find the code and pre-trained models useful for your research, please consider citing our paper. :blush:

@inproceedings{liu2025frequency,
  title={Frequency Domain-Based Diffusion Model for Unpaired Image Dehazing},
  author={Liu, Chengxu and Qi, Lu and Pan, Jinshan and Qian, Xueming and Yang, Ming-Hsuan},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2025}
}