DARD: Zero-Shot Degradation-Aware Retinex-Guided Diffusion for Low-Light Image Enhancement

IEEE Transactions on Circuits and Systems for Video Technology, 2026

Wenjie Cai, Yuezhe Yang, Jianyang Xia, Xingbo Dong^†, Zhe Jin^†

^†Corresponding author

Anhui University

Paper

Abstract

Existing diffusion-based low-light image enhancement methods often require paired training data or lack reliable scene constraints in zero-shot settings, leading to structural inconsistencies and color drift. We present DARD, a zero-shot degradation-aware Retinex-guided diffusion framework. DARD extracts image-specific physical priors with test-time Retinex decomposition, combines them with diffusion predictions using timestep-adaptive frequency fusion, and refines the reverse process with physical-consistency and CLIP-based semantic guidance. Evaluations on real-world low-light benchmarks show strong restoration performance. Enhanced images also improve downstream semantic segmentation, including a 28.10% relative mIoU gain over AGLLDiff on LOLv1.

Highlights

Calibrate degradation-aware Retinex priors for each low-light input.
Fuse physical priors with diffusion predictions across timesteps.
Improve restoration and downstream segmentation without paired training.

Method

DARD architecture with Retinex decomposition, frequency fusion, and reverse refinement

Overview of the test-time decomposer, timestep-adaptive frequency fusion, and guided reverse diffusion refinement.

Results

Quantitative Results

Method	LOLv1			LOLv2-real
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
AGLLDiff	19.83	0.814	0.194	19.99	0.844	0.192
DARD	21.14	0.835	0.143	22.18	0.877	0.135

Zero-shot low-light enhancement comparison on LOLv1 and LOLv2-real.

Dataset	SegFormer + AGLLDiff mIoU	SegFormer + DARD mIoU	Relative improvement
LOLv1	21.32	27.31	+28.10%
LOLv2-real	28.09	35.13	+25.06%

Downstream semantic segmentation using enhanced images.

Qualitative Results

Representative low-light inputs and DARD outputs

Representative enhancement results across real-world low-light scenes.

Comparison with low-light image enhancement methods

Visual comparison on the LOLv1 and LOLv2-real datasets.

Downstream semantic segmentation comparison

Semantic segmentation results on low-light scenes.

Citation

@article{cai2026dard,
  title={DARD: Zero-Shot Degradation-Aware Retinex-Guided Diffusion for Low-Light Image Enhancement},
  author={Cai, Wenjie and Yang, Yuezhe and Xia, Jianyang and Dong, Xingbo and Jin, Zhe},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2026},
  url={https://ieeexplore.ieee.org/abstract/document/11701592}
}