This repository contains the official implementation of the ICCV 2025 paper " Balancing Task-invariant Interaction and Task-specific Adaptation for Unified Image Fusion" .

• python 3.9
• torch 1.12.1
• cudatoolkit 11.3
• torchvision 0.13.1
• numpy 1.23.5
• opencv-python 4.7.0
• ...

The datasets is constructed following TC-MoA. (Or you can download from google drive).

The data structure is like this:

dataset_name
├── subdir1
│  ├── xxx.png
│  ├── ...
├── subdir2
│  ├── xxx.png
│  ├── ...

For training:

• LLVIP
• SCIE
• RealMFF, MFI-WHU

For testing:

• LLVIP
• MEFB
• MFFB (Lytro, MFFW, MFI-WHU)

TITA results are available on google drive.

Download the pretrained checkpoint from google drive, and put it under ./logs/mixed/models/. Change the dataset name and path in test.py. And the results can be found in ./results_tita/.

bash test.bash

Change the dataset name and path in config.json and run:

bash train.bash

@inproceedings{hu2025balancing,
  title={Balancing Task-invariant Interaction and Task-specific Adaptation for Unified Image Fusion},
  author={Hu, Xingyu and Jiang, Junjun and Wang, Chenyang and Jiang, Kui and Liu, Xianming and Ma, Jiayi},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2025}
}

This code is mainly built upon SwinFusion and FAMO. Thanks for their excellent work!