[CVPR 2026] FusionRegister

This is the official PyTorch implementation of the paper FusionRegister: Every Infrared and Visible Image Fusion Deserves Registration [paper] [arxiv].

Contents

• Requirements and dependencies
• Dataset
• Models
• Test
• Train
• Citation
• Contact
• Acknowledgement

Requirements and dependencies

The model is built in PyTorch 2.1.0 (Python3.9.19, CUDA12.8).

For installing, follow these intructions:

conda create -n pytorch python=3.9
conda activate pytorch
conda install pytorch==2.1.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
pip install requirements.txt

Install warmup scheduler

cd pytorch-gradual-warmup-lr
python setup.py install
cd ..

Dataset

Registration dataset

1. Download dataset from the Registered dataset .

2. Unzip files dataset folder.

3. Make Infrared and Visible image fusion dataset structure be:

Registration Dataset
├─ dataset
│ ├─ ir
│ │ ├─ img1.png
│ │ ├─ ......
│ ├─ vi
│ │ ├─ img1.png
│ │ ├─ ......
├─ ......

Models

Pre-trained models can be downloaded from google drive.

• fr.pth: trained for MMDRFuse (ACMM 24) (Main Version).
• fr_hcl.pth: trained for HCLFuse (NeurIPS 25) (Additional Version).
• fr_s4.pth: trained for S4Fusion (TIP 25) (Additional Version).

Test MMDRFuse Version (Main Version)

1. Clone this github repo

git clone https://github.com/bociic/FusionRegister.git
cd FusionRegister

2. Prepare testing dataset and modify "input_dir", "target_dir", and "weights" in ./test.py
3. Run test

python test.py --save_result #test MMDRFuse (Main Version)

4. The original fused result are saved in ./fused_results. The registered result are saved in ./registered_results

Test S4Fusion Version

1. Download the pretrained weight of image fusion from S4Fusion, and download the pretrained weight of FusionRegister(S4-Version). Put them under ./checkpoints.
2. Prepare testing dataset and modify "input_dir", "register_weights", and "fusion_weights" in ./test_s4.py
3. Run test

python test_s4.py --save_result #test S4Fusion (TIP 2025)

4. The original fused result of S4Fusion are saved in ./fused_results. The registered result of S4Fusion are saved in ./registered_results

Test HCLFuse Version

1. Download the pretrained weight of image fusion from HCLFuse, and download the pretrained weight of FusionRegister(HCL-Version). Put them under ./checkpoints.
2. Prepare testing dataset and modify "input_dir", "register_weights", and "fusion_weights" in ./test_hcl.py
3. Run test

python test_hcl.py --save_result #test HCLFuse(NeurIPS 2025) 

4. The original fused result of HCLFuse are saved in ./fused_results. The registered result of HCLFuse are saved in ./registered_results

Train

1. Clone this github repo

git clone https://github.com/ChengxuLiu/MISCFilter.git
cd MISCFilter
pip install requirements.txt

2. Prepare training dataset and modify "input_dir" and "fusion_weights" in ./train.py. If you want to finetune from the pretrained weight you can modify "pretrained_weights" in ./train.py
3. If you want to train your own version, modify Load the Fusion Net and Get the Original Fused Image in Line 106 and Line 175 in train.py
4. Run training

python train.py

4. The models are saved in ./experiments

Citation

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

@misc{bian2026fusionregisterinfraredvisibleimage,
      title={FusionRegister: Every Infrared and Visible Image Fusion Deserves Registration}, 
      author={Congcong Bian and Haolong Ma and Hui Li and Zhongwei Shen and Xiaoqing Luo and Xiaoning Song and Xiao-Jun Wu},
      year={2026},
      eprint={2603.07667},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.07667}, 
}

Contact

If you meet any problems, please describe them in issues or contact:

• Congcong Bian: bociic_jnu_cv@163.com

Acknowledgement

The code of FusionRegister is built upon MISCFilter,and we express our gratitude to this awesome projects.