Highlight What You Want

Codes for Highlight What You Want: Weakly-Supervised Instance-Level Controllable Infrared-Visible Image Fusion. (ICCV 2025) Paper

---

---

1.Create Envrionment

• Create Conda Environment

conda create -n ris_fusion_env python=3.10
conda activate ris_fusion_env

• Install Dependencies

pip install -r requirements.yml

---

2.Pre-trained Weights

The Pre-trained weights for fusion is at weights.

Please put the 'best_model.pth' in 'RISW-STAGE2-OS/CRIS/exp/refcoco/CRIS_R50' and the 'model_stage2.pth' in 'RISW-STAGE2-OS/'

---

3.Testing

For generate fusion image, please run

python ./RISW-STAGE2-OS/test_one_img.py

Result will be saved in the './RISW-STAGE2-OS/test_rusults/' folder.

---

4.Training

Training Set for Stage II is at Train.

RISW-STAGE1-OS

Please prepare the training data at

data/ 
	IVT_train/
    	ir/
            1.png
            2.png
       		 ...
        vi/
            1.png
            2.png
            ...
        text/
            1.txt
            2.txt
            ... 	

Run the code

python train_stage1.py

---

RISW-STAGE2-OS

Please prepare the training data at

dataset/ 
	IVT_final/
    	IVT_ir/
            1.png
            2.png
       		 ...
        IVT_vi/
            1.png
            2.png
            ...
        IVT_text/
            1.txt
            2.txt
            ... 	

After that, run the following command:

python dataprocess.py
python train_fused.py

---

5.Fusion Results

1.Qualitative comparison of various fusion models.

---

2.Quantitative comparison of various fusion models across four test sets.

---

6.Citation

@InProceedings{Wang_2025_ICCV,
    author    = {Wang, Zeyu and Zhang, Jizheng and Song, Haiyu and Ge, Mingyu and Wang, Jiayu and Duan, Haoran},
    title     = {Highlight What You Want: Weakly-Supervised Instance-Level Controllable Infrared-Visible Image Fusion},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {12637-12647}
}