[2025-10-13] The arXiv version of the paper is available here.
[2025-10-7] The code and dataset are available.
[2025-9-30] Our paper has been accepted by NeurIPS 2025, and the code will be released soon.
The preview of our dataset is as follows.
The datasets containing clean and degraded infrared images are provided in:
For quick start, first, you can create a new environment:
conda create -n PPFN python=3.8
conda activate PPFN
pip install -r requirements.txt
Then download the pretrained model weights.
You can run the following command:
CUDA_VISIBLE_DEVICES=0 python demo.py --dataset_dir ./example/{normal, hard} --output example/enhanced --weights put/weight/path/here
You should put the dataset in the "datasets" fold in the following form.
HM-TIR ROOT
├── datasets
| ├── HM-TIR
| | ├── imgs # all clean infrared images
| | ├── composited # test composited degraded infrared images
| | | ├── normal # sub test set in normal degradation
| | | └── hard # sub test set in hard degradation
| | ├── single # single degraded infrared images
| | | ├── blur # sub test set in blur degradation
| | | ├── contrast # sub test set in contrast degradation
| | | └── noise # sub test set in noise degradation
| | ├── test # test clean infrared images
| | └── meta # meta data, includes: train.txt, test.txt
| └── ...
Download the HM-TIR dataset, then run the command:
python train.py --dataset_name HM-TIR --dataset_dir datasets
For composited degradation scenario, run the following command:
python evaluation.py --dataset_name HM-TIR --dataset_dir datasets --deg_scenario composited --subset {normal, hard} --output results --weights put/weight/path/here
For single degradation scenario, run the following command:
python evaluation.py --dataset_name HM-TIR --dataset_dir datasets --deg_scenario single --subset {contrast, blur, noise} --output results --weights put/weight/path/here
For real-world dataset evaluation, we select 70 images form Iray. You can download the images we selected, then run the following command:
python inference.py --dataset_name Iray_select --dataset_dir datasets --output results --cal_iqa True --weights put/weight/path/here
You can download some results in here
If you want to generate new degraded TIR Images with new degradation levels or types, first, changing the "create_corrupt.py" file about degradation setting (lines 18-68), then, run the following command:
python create_corrupt.py --dataset_name HM-TIR --dataset_dir datasets --save_dir datasets/corrputed
If you have any other questions about the code and dataset, please email to Zihang Chen or Jinyuan Liu.
Please cite us if our work is useful for your research.
@inproceedings{liu2025enhancing,
title = {Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark},
author = {Jinyuan Liu and Zihang Chen and Zhu Liu and Zhiying Jiang and Long Ma and Xin Fan and Risheng Liu},
booktitle = {The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year = {2025},
url = {https://openreview.net/forum?id=yEddfz9SgJ}
}
Our core codes are based on MIOIR and LLFormer, thanks for their contribution.