Dereflection Any Image with Diffusion Priors and Diversified Data

AAAI 2026

Project Page | Paper | | | |

📖Dereflection Any Image with Diffusion Priors and Diversified Data

Jichen Hu^1*, Chen Yang^1*, Zanwei Zhou¹, Jiemin Fang^2†, Qi Tian², Wei Shen^1✉†,
¹MoE Key Lab of Artificial Intelligence, AI Institute, School of Computer Science, Shanghai Jiao Tong University   ²Huawei Inc.  
^*Equal contribution.   ^†Project lead.   ^✉Corresponding author.

📝 Todo

• Release inference code
• Release pretrained model weights
• Release project page
• Release paper
• Release dataset
• Release training code

🚀Setup

Environment

Dereflection Any Image is tested with CUDA 11.8 and python 3.9. All the required packages are listed in requirements.txt. You can install them with

pip install -r requirements.txt

Weights

Our scripts will automatically download the pretrained weights from Huggingface.

You can also download the weights by Google Drive or Huggingface

💪Usage

Inference

Put images in the "input" dir, and run:

python run.py --input_dir ./input/ --result_dir ./result/ --concat_dir ./concat/ 

or use script run.sh directly.

Gradio Demo

python demo.py

🏋️‍♂️ Training

Data preparation

You can download the training and test data by Huggingface or run the download_merge.sh, the dir structure is like:

dataset_root/
├── train/
│   ├── nature/
│   ├── real/
│   ├── DRR/
│   ├── Syn/
│   ├── DRR_dual/
│   ├── Syn_dual/
│
└── test/
    ├── real20_420/
    ├── Nature/
    ├── SIR2/
    ├── DRR-S/
    ├── DRR-C/

Run create_jsonl.py to create jsonl files for training.

Progressive training

# change to your local path before running

# foundamental training
sh train.sh

# reflection-invariant fine-tuning
sh train_dual.sh

# train the CLD
sh train_decoder.sh

Test

python run_test.py
python metrics.py -m /path/to/output

🌏 Citation

If you find Derefelection Any Image useful for your work please cite:

@misc{hu2025dereflection,
    title={Dereflection Any Image with Diffusion Priors and Diversified Data},
    author={Jichen Hu and Chen Yang and Zanwei Zhou and Jiemin Fang and Xiaokang Yang and Qi Tian and Wei Shen},
    year={2025},
    eprint={2503.17347},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}