• 2025/05 The paper is accepted by Pattern Recognition: https://doi.org/10.1016/j.patcog.2025.111816
• We have released our code.

It is recommended to install the environment with environment.yaml.

conda env create --file=environment.yaml

Download KAIST dataset from https://github.com/SoonminHwang/rgbt-ped-detection

Download FLIRv1 dataset from https://www.flir.com/oem/adas/adas-dataset-form/

Download VEDAI dataset from https://downloads.greyc.fr/vedai/

We adopt the official dataset split in our experiments.

VQGAN can be downloaded from https://ommer-lab.com/files/latent-diffusion/vq-f8.zip (Other GAN models can be downloaded from https://github.com/CompVis/latent-diffusion).

TeVNet and PID heckpoints can be found in HuggingFace.

Use the shellscript to evaluate. indir is the input directory of visible RGB images, outdir is the output directory of translated infrared images, config is the chosen config in configs/latent-diffusion/config.yaml. We prepare some RGB images in dataset/KAIST for quick evaluation.

bash run_test_kaist512_vqf8.sh

Prepare corresponding RGB and infrared images with same names in two directories.

cd TeVNet
bash shell/train.sh

To accelerate training, we recommend using our pretrained model.

bash shell/run_train_kaist512_vqf8.sh

Our code is built upon LDM and HADAR. We thank the authors for their excellent work.

If you find this work is helpful in your research, please consider citing our paper:

@article{mao2026pid,
  title={PID: physics-informed diffusion model for infrared image generation},
  author={Mao, Fangyuan and Mei, Jilin and Lu, Shun and Liu, Fuyang and Chen, Liang and Zhao, Fangzhou and Hu, Yu},
  journal={Pattern Recognition},
  volume={169},
  pages={111816},
  year={2026},
  publisher={Elsevier}
}

If you have any question, feel free to contact [email protected].