Noise2DiffusionEnhanced : A Pretrained Ready-to-use TIR Denoising Transformer

Our approach uses diffusion-based image enhancement and realistic TIR image degradation to generate image pairs for supervised learning (a) and leverages remarkable visual quality of diffusion models (c) without suffering from hallucinations (d-e).

Overview

Files

• model_best.pth is the pretrained-weight file. It can only be used with the associated project codebase, where is described how to insert it to run inference or fine-tune the weights.

Citation

@article{hazebrouck_self-supervised_2026,
    title = {Self-supervised Diffusion-guided Hallucination-free Thermal Infrared Image Denoising},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    author = {Hazebrouck, Félix and Schock-Schmidtke, Alexander and Stuhrmann, Norbert and Fottner, Johannes and Teutsch, Michael},
    year = {2026},
    pages = {7091--7101},
}

License and Copyrights

• The model architecture is a U-Former, introduced in [1]. The original code was adapted for our project and the modified version along with all license-notices with respect to the original source is available on our project codebase and is licensed under the MIT license. As we publish no code parts here, and because the code of [1] only indirectly contributed to those weights, we cite them only for transparency.
• The model was trained with our HDRT-TIR-DE dataset, which is licensed under the CC-BY-NC-SA-4.0 license. The HDRT-TIR-DE dataset itself is a modified version of the HDRT dataset introduced in [2], but because the HDRT-TIR-DE already comply with the licensing of the HDRT, as explicitly stated on the HDRT-TIR-DE repository and only indirectly contributed to the weights, only the license from the HDRT-TIR-DE dataset applies here.
  According to the license of the HDRT-TIR-DE Dataset (CC-BY-NC-SA-4.0), any "adapted material" must be licensed under the same or a compatible license, therefore including these pre-trained weights.
• The degradation model used to synthesize the noisy counterparts for the clean images from the HDRT-TIR-DE dataset for the supervised training of the UFormer is adapted from the noise model introduced in [3]. The original code was adapted for our project and the modified version along with all license-notices with respect to the original source is available on our project codebase and is licensed under the MIT license. As we publish no code parts here, and because the code of [3] only indirectly contributed to those weights, we cite them only for transparency.

References