Noise2DiffusionEnhanced

 

<div align="center"> <table> <tr> <td align="center" height="60"> <a href="https://openaccess.thecvf.com/content/CVPR2026W/PBVS/html/Hazebrouck_Self-supervised_Diffusion-guided_Hallucination-free_Thermal_Infrared_Image_Denoising_CVPRW_2026_paper.html"> <img src="./README_figures/button_paper.png" height="50"/> </a> </td> <td align="center" height="60"> <a href="https://github.com/HensoldtOptronicsCV/Noise2DiffusionEnhanced"> <img src="./README_figures/button_github.png" height="50"/> </a> </td> <td align="center" height="60"> <a href="https://huggingface.co/collections/SachyGuy/diffusion-guided-hallucination-free-tir-image-denoising"> <img src="./README_figures/button_huggingface.png" height="50"/> </a> </td> </tr> </table> </div>

 

---

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

This is part of the official repository of the paper Self-supervised Diffusion-guided Hallucination-free Thermal Infrared Image Denoising, 2026.
 

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

The Noise2DiffusionEnhanced is part of the work described in "Self-supervised Diffusion-guided Hallucination-free Thermal Infrared Image Denoising", F. Hazebrouck, A. Schock-Schmidtke, N. Stuhrmann, J. Fottner, M. Teutsch (2026). The paper is available here.

The Noise2DiffusionEnhanced is a pretrained Uformer architecture [1] for TIR sensor noise denoising. It was trained on the HDRT-TIR-diffusion-enhanced dataset. The pretrained weights published here can be inserted in the GitHub Project for inference or fine-tuning.

This pretrained model should contribute to filling the current lack of reference denoising models for Thermal Infrared (TIR) single-image denoising, for direct use as well as for related research.

 

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

If you use the Noise2DiffusionEnhanced, please cite our work:

@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 weights were obtained using different code-compounds, each with its proper license:

• 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.

Therefore, the pre-trained weights published in this repository inherit the CC-BY-NC-SA-4.0 license from the HDRT-TIR-DE dataset. According to this license, we include the original dataset authors, a copyright notice, a link to original dataset, a link to license, and a statement of modifications.

Authors:
© Félix Hazebrouck, Alexander Schock-Schmidtke, Norbert Stuhrmann, Johannes Fottner, Michael Teutsch

Original dataset:
HDRT-TIR-DE dataset, available at https://huggingface.co/datasets/SachyGuy/HDRT-TIR-DE, introduced in Self-supervised Diffusion-guided Hallucination-free Thermal Infrared Image Denoising by Hazebrouck et al., 2026.

License:
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
https://spdx.org/licenses/CC-BY-NC-SA-4.0

Modifications:
We encoded the statistical information from the HDRT-TIR-DE dataset in the hereby published pretrained neural-network architecture weights through a process of machine learning.

These pretrained weights are distributed under the same license (CC-BY-NC-SA-4.0).

References

<a id="1">[1]</a> Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. CVPR, 2022

<a id="2">[2]</a> Jingchao Peng, Thomas Bashford-Rogers, Francesco Banterle, Haitao Zhao and Kurt Debattista. HDRT: A large-scale dataset for infrared-guided HDR imaging. Elsevier Information Fusion, 120, 2025

<a id="3">[3]</a> Lijing Cai, Xiangyu Dong, Kailai Zhou and Xun Cao. Exploring video denoising in thermal infrared imaging: Physics-inspired noise generator, dataset, and model, IEEE 2022