[CVPR 2026] ZeroIDIR: Zero-Reference Illumination Degradation Image Restoration with Perturbed Consistency Diffusion Models

Hai Jiang¹, Zhen Liu², Yinjie Lei³, Songchen Han¹, Bing Zeng², Shuaicheng Liu²

1.School of Aeronautics and Astronautics, Sichuan University

2.University of Electronic Science and Technology of China,

3.College of Electronics and Information Engineering, Sichuan University

Overall pipeline

Dependencies

pip install -r requirements.txt

Download the raw training and evaluation datasets

LLIE datasets

BIE datasets

MSEC datasets

Real-world datasets

Pre-trained Models

You can download our pre-trained model from [OneDrive] and [Baidu Yun (extracted code:)]

How to train?

You need to modify dataset/dataloader.py slightly for your environment, and then

accelerate launch train.py  

How to test?

python inference.py

Visual comparison

Citation

If you use this code or ideas from the paper for your research, please cite our paper:

Acknowledgement

Part of the code is adapted from the previous work: denoising-diffusion-pytorch. We thank all the authors for their contributions.