Efficient Rectified Flow for Image Fusion

Efficient Rectified Flow for Image Fusion [NeurIPS 2025]

<div align=center> <img src="https://github.com/zirui0625/RFfusion/blob/main/figures/pipeline.png" width="90%"> </div>

Updates

[2025-9-25] Our paper has been accepted by NeurIPS 2025. You can find our paper here.

Environment

# create virtual environment
conda create -n RFfusion python=3.8
conda activate RFfusion
# install requirements
pip install -r requirements.txt

Test

Download the Rectified Flow checkpoint from here, we use 'checkpoint12.pth' for sampling, and put it in './model'. Our pretrained VAE model can be found in here, also put it in './model'. You can test our method through

python sample.py

Train

Stage I

You can download the training data from LLVIP, MSRS, and place it in ./data. You can download the initial VAE model (f=4, VQ) from LDM and place it in ./model.

CUDA_VISIBLE_DEVICES=0 python train_vae.py -b ./vae_config.yaml -t -r ./model/model.ckpt --gpus 0,

Stage II

You can download the training data from MSRS, and place it in ./data.

CUDA_VISIBLE_DEVICES=0 python train_fusion.py,

Results

<div align=center> <img src="https://github.com/zirui0625/RFfusion/blob/main/figures/result1.png" width="100%"> </div> <div align=center> <img src="https://github.com/zirui0625/RFfusion/blob/main/figures/result2.png" width="100%"> </div>

Citation

@article{wang2025efficient,
  title={Efficient Rectified Flow for Image Fusion},
  author={Wang, Zirui and Zhang, Jiayi and Guan, Tianwei and Zhou, Yuhan and Li, Xingyuan and Dong, Minjing and Liu, Jinyuan},
  journal={arXiv preprint arXiv:2509.16549},
  year={2025}
}

Contact

If you have any questions, feel free to contact me through <code style="background-color: #f0f0f0;">ziruiwang0625@gmail.com</code>。

Acknowledgement

Our codes are based on Rectified Flow, LDM, DDFM, thanks for their contribution.