[CVPR 2025] Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing 🔗

This repository presents the implementation of the paper

Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing
Ruiyi Wang, Yushuo Zheng, Zicheng Zhang, Chunyi Li, Shuaicheng Liu, Guangtao Zhai, Xiaohong Liu
The IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR), 2025

We present a novel hazing-dehazing pipeline consisting of a Realistic Hazy Image Generation framework (HazeGen) and a Diffusion-based Dehazing framework (DiffDehaze).

🛠️ Setup

📦 Repository

Clone the repository (requires git):

git clone https://github.com/ruiyi-w/Learning-Hazing-to-Dehazing.git
cd Learning-Hazing-to-Dehazing

💻 Dependencies

Using Conda. After the installation, create the environment and install dependencies into it:

conda create -n env_name python=3.10
conda activate env_name
pip install -r requirements.txt

Keep the environment activated before running the inference script. Activate the environment again after restarting the terminal session.

🏃 Testing

📷 Prepare input images

Place your images in the inputs/ directory. To set a different source directory, you can edit configuration files in configs/inference/.

⬇ Download Checkpoints

Download pre-trained models and place them to folder weights/, but you can always edit configuration files in configs/inference/.

Model Name	Description	Link
v2-1_512-ema-pruned.ckpt	Pretrained Stable Diffusion v2.1 from stabilityai, providing generative priors	download
stage1.pt	IRControlNet trained for HazeGen	download
stage2.pt	IRControlNet trained for DiffDehaze	download

🚀 Run inference

To perform dehazing with standard spaced sampler, please run

python inference_stage2.py --config configs/inference/stage2.yaml

By default, results will be saved to outputs/. Enjoy!

To perform dehazing with AccSamp sampler, please run

python inference_accsamp.py --config configs/inference/stage2_accsamp.yaml

To generate realisitic hazy images with HazeGen, please run

python inference_stage1.py --config configs/inference/stage1.yaml

To use a different hyperparameter settings, e.g., $\tau$ and $\omega$, please edit the corresponding .yaml configuration file.

🏋️ Training

1. Training data preparation.

   • The training of HazeGen requires real-world hazy data from the URHI split of RESIDE dataset and the synthetic hazy data from RIDCP.
   • To train DiffDehaze, you need to generate realistic hazy data from HazeGen based on clean images, e.g., the clean images from the OTS split of RESIDE dataset, using the inference script above.

2. Accelerate configuration. The training is supported by the huggingface Accelerate library. Before running training scripts, create and save a configuration file to help Accelerate correctly set up training based on your setup by running

accelerate config

3. Fill in the training configuration files in configs/train/ with appropriate values, especially for the paths to the training data. Please find specific instructions there.
4. Start training! To train stage1 HazeGen model, run

accelerate launch train_stage1.py --config configs/train/stage1.yaml

5. To train stage2 DiffDehaze model, run

accelerate launch train_stage2.py --config configs/train/stage2.yaml

✏️ Acknowledgment

A large part of the implementation is based on DiffBIR. We sincerely appreciate their wonderful work.

🎓 Citation

If you find our work useful, please consider cite our paper:

@misc{wang2025learninghazingdehazingrealistic,
      title={Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing}, 
      author={Ruiyi Wang and Yushuo Zheng and Zicheng Zhang and Chunyi Li and Shuaicheng Liu and Guangtao Zhai and Xiaohong Liu},
      year={2025},
      eprint={2503.19262},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.19262}, 
}

🎫 License

This work is licensed under the Apache License, Version 2.0 (as defined in the LICENSE).

By downloading and using the code and models you agree to the terms in the LICENSE.