Toward Real-world Infrared Image Super-Resolution: A Unified Autoregressive Framework and Benchmark Dataset

Yang Zou, Jun Ma, Zhidong Jiao, Xingyuan Li, Zhiying Jiang, and Jinyuan Liu, "Toward Real-world Infrared Image Super-Resolution: A Unified Autoregressive Framework and Benchmark Dataset", CVPR 2026 Highlight

:rocket: Updates

[2026-3-10] Our training code and inference code is now available.🎊🎊🎊

[2026-3-6] You can find our paper here. ⭐️⭐️⭐️

[2026-3-4] Our dataset is now available.🔥🔥🔥

[2026-2-21] Our paper has been accepted by CVPR 2026. The code and dataset have been officially released.🎉🎉🎉

💾 FLIR-IISR Dataset 💾

:open_book: Dataset Details

Download

Composition ($1457$ pairs)

• Scene labels ($12$ categories):

  • person ($309$), bicycle ($22$), motorcycle ($27$), tricycle ($13$), car ($234$), bus ($5$) plane ($54$), statue ($157$), regular object ($248$), building ($706$), road ($132$), and complex scene ($401$).

• Degradation labels:

  • Optical blur ($1305$); Motion blur ($152$).

• Total number of image pairs: $1457$

• Image size: $1024 \times 768$

Preview

---

Scene labels:

Degradation labels:

---

📦 Real-IISR 📦

⚙️ Dependencies

git clone https://github.com/JZD151/Real-IISR.git
cd Real-IISR

conda create -n Real-IISR python=3.10
conda activate Real-IISR
pip install -r requirements.txt
pip install flash_attn-2.7.4.post1 --no-build-isolation

🔧 Training

1. Download the pretrained VQVAE and VARSR models, and place them in the ./checkpoints directory.
2. Download the FLIR-IISR dataset and extract it.

python train.py --batch_size 4 --ep 20 --fp16 1 --tblr 5e-5 --alng 1e-4 --wpe 0.01 --fuse 0 --exp_name Real-IISR

🔨 Testing

Note: We provide several sample inputs for easy inference.

1. Download the pretrained model from 百度网盘 / HuggingFace , and place it in the ./checkpoints directory.

python test.py

📎 Citation

@inproceedings{zou2026toward,
  title={Toward real-world infrared image super-resolution: A unified autoregressive framework and benchmark dataset},
  author={Zou, Yang and Ma, Jun and Jiao, Zhidong and Li, Xingyuan and Jiang, Zhiying and Liu, Jinyuan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={16365--16375},
  year={2026}
}

📫 Contact

If you have any questions, feel free to contact us through archerv2@mail.nwpu.edu.cn.

💡 Acknowledgements

Our codes are based on VAR, VARSR, thanks for their contribution.