Organizations: Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen 518083, China · Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, School of Artificial Intelligence, Anhui University, Hefei 230601, China · State Key Laboratory of Opto-Electronic Information Acquisition and Protection Technology, the Anhui Provincial Key Laboratory of Secure Artificial Intelligence, Anhui Provincial International Joint Research Center for Advanced Technology in Medical Imaging, and the School of Artificial Intelligence, Anhui University, Hefei, China
Existing diffusion-based enhancement methods provide strong generative capability for low-light image enhancement (LLIE), yet they either rely on paired supervision or lack reliable scene constraints in zero-shot settings, often leading to structural inconsistency and color drift. Motivated by conventional Retinex models, which offer physically interpretable priors that can serve as reliable scene constraints yet struggle with mixed degradations in real-world scenarios, we propose DARD, a zero-shot Degradation-Aware Retinex-guided Diffusion framework for LLIE. DARD first extracts image-specific physical priors from the degraded input through a test-time degradation-aware Retinex decomposition, thereby providing reliable structural guidance for zero-shot restoration. It then injects these priors into reverse diffusion through a timestep-adaptive frequency fusion strategy to balance structural anchoring and detail generation. Finally, a guided reverse refinement process with physical consistency and Contrastive Language-Image Pre-training (CLIP)-based semantic guidance is introduced to suppress structural artifacts and semantic drift during sampling. Extensive experiments show that DARD achieves strong distortion and perceptual performance and consistently outperforms existing zero-shot baselines across multiple real-world low-light benchmarks. To further validate the practical utility of our method for downstream applications, we evaluated its impact on semantic segmentation. Experiments demonstrate that images enhanced by DARD achieve a 28.10% relative improvement in mIoU over AGLLDiff.
Figures & tables
Fig. 1: Conceptual comparison of (a) Retinex-based, (b) diffusion-based, and (c) DARD paradigms. Unlike existing methods, DARD explicitly determines where the physical prior is reliable and when it should guide reverse diffusion.
Fig. 2: Representative results of DARD on diverse real-world low-light scenes, including (a) dusk scenes, (b) natural landscapes, (c) architecture, (d)indoor objects, (e) human faces and (f) animals. In each example, the left half shows the low-light input and the right half shows the enhanced result.
Fig. 3: Overall Architecture of the proposed DARD. Our method consists of three parts: (a) Test-Time Degradation-Aware Retinex Decomposer, (b) Timestep-Adaptive Frequency Fusion (TAFF), and (c) Guided Reverse Diffusion Refinement (GRDR). Specifically, the Decomposer extracts physical priors via test-time optimization. TAFF then fuses them with diffusion predictions in the frequency domain based on σt , and GRDR iteratively refines the intermediate x~0 using physics and CLIP guidance.
Fig. 4: Architecture of the Test-Time Degradation-Aware Retinex Prior Calibration module. Starting from the frozen pretrained Retinex decomposition, Φdeg characterizes image-specific residual degradation and Φconf estimates the spatial reliability of the physical prior.
Method
Type
LOLv1
LOLv2-real
LSRW-Huawei
LSRW-Nikon
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
KinD(MM’19) [ 22 ]
S
16.55
0.429
0.405
15.11
0.402
0.426
17.35
0.530
0.544
14.69
0.471
0.469
URetinexNet(CVPR’22) [ 24 ]
S
19.84
0.824
0.121
21.09
0.858
0.144
19.14
0.553
0.419
16.53
0.452
0.264
GSAD(NeurIPS’23) [ 38 ]
S
22.98
0.851
0.103
20.19
0.847
0.178
20.43
0.583
0.407
17.36
0.506
0.222
DiffLL(Siggraph Asia’23) [ 11 ]
S
26.34
0.845
0.217
28.86
0.876
0.207
18.06
0.417
0.482
15.94
0.359
0.402
CUGD(TCSVT’25) [ 34 ]
S
22.72
0.865
0.111
23.17
0.882
0.112
19.64
0.588
0.327
16.33
0.468
0.211
TABLE I: Quantitative comparison on the LOLv1 [ 21 ] , LOLv2-real [ 36 ] , and LSRW [ 37 ] . “S” and “U” represent “Supervised” and “Unsupervised” methods, respectively. The best results of “S” and “U” are marked in blue and orange , respectively.
Method
DICM
MEF
LIME
VV
NPE
Q-Align ↑
MUSIQ ↑
LOE ↓
Q-Align ↑
MUSIQ ↑
LOE ↓
Q-Align ↑
MUSIQ ↑
LOE ↓
Q-Align ↑
MUSIQ ↑
LOE ↓
Q-Align ↑
MUSIQ ↑
LOE ↓
KinD [ 22 ]
2.46
49.38
317.45
2.96
49.84
246.35
2.46
48.69
218.36
2.57
32.47
264.32
1.64
48.74
231.57
URetinexNet [ 24 ]
3.56
62.28
222.73
3.63
65.96
141.84
3.36
59.93
139.43
3.01
40.46
125.36
3.47
60.11
198.76
GSAD [ 38 ]
3.81
60.36
194.46
3.62
61.76
126.83
3.28
58.64
126.18
3.22
47.95
122.63
3.56
63.49
122.63
DiffLL [ 11 ]
3.43
58.27
191.31
3.27
62.58
142.13
3.08
57.64
151.82
3.59
63.59
164.82
3.27
65.63
171.62
CUGD [ 34 ]
3.90
61.44
99.03
3.57
63.08
99.74
3.26
60.37
120.41
3.19
61.53
129.31
3.47
67.06
139.12
TABLE II: Quantitative comparison on the DICM [ 41 ] , MEF [ 42 ] , LIME [ 3 ] , VV [ 43 ] , and NPE [ 44 ] datasets. The best results of supervised and unsupervised methods are marked in blue and orange , respectively. Note that we report the quality score for the Q-Align metric.
Fig. 5: Qualitative comparison with the state-of-the-art low-light image enhancement methods on the LOLv1 [ 21 ] and LOLv2-real [ 36 ] datasets.
Fig. 6: Qualitative comparison with the state-of-the-art low-light image enhancement methods on the LSRW [ 37 ] dataset.
TABLE III: Quantitative comparison of ablation experiments on the LOLv1 dataset.
λphys
λpseudo
λstruct
λcontent
PSNR ↑
SSIM ↑
LPIPS ↓
0
1
1
1
16.51
0.432
0.257
1
0
1
1
16.24
0.468
0.261
1
1
0
1
16.89
0.486
0.243
1
1
1
0
16.78
0.481
0.315
0.5
0.5
0.5
0.5
16.06
0.416
0.324
1
1
1
1
16.79
0.483
0.257
TABLE IV: Ablation study of the guidance weights in GRDR on the LSRW-Nikon dataset. The last row is the default setting.
Physics Guidance
CLIP Guidance
PSNR ↑
SSIM ↑
LPIPS ↓
20.31
0.816
0.174
✓
20.48
0.828
0.154
✓
20.53
0.819
0.162
✓
✓
21.14
0.835
0.143
TABLE V: Ablation study on multimodal guidance on the LOLv1 dataset.
Fig. 8: Visualization of ablation results.
Steps
PSNR ↑
SSIM ↑
LPIPS ↓
10
19.66
0.805
0.186
50
20.29
0.819
0.172
200
20.96
0.824
0.141
100
21.14
0.835
0.143
TABLE VI: Ablation analysis of the diffusion steps on the LOLv1 dataset.
Prompt Setting
PSNR ↑
SSIM ↑
LPIPS ↓
A
20.68
0.826
0.164
B
20.82
0.829
0.152
C (Ours)
21.14
0.835
0.143
TABLE VII: Ablation analysis of different CLIP prompt on the LOLv1 dataset.
Method
Steps
Params (M)
FLOPs (T)
Memory (GB)
Time (s)
AGLLDiff
10
553
109
3.07
2.01
QuadPrior
10
1314
6
5.68
0.69
FourierDiff
50
548
110
2.82
2.19
LLIEDiff
999
1066
1971
6.87
65.05
Ours
100
704
222
3.27
4.46
TABLE VIII: Computational efficiency comparison of representative LLIE methods.
Fig. 9: Qualitative comparison of semantic segmentation results on low-light scenes. The baseline method, SegFormer [ 51 ] , uses the low-light image as input, while SegFormer [ 51 ] +Ours uses the enhanced image as input.
Method
LOLv1
LOLv2-real
aAcc ↑
mIoU ↑
mAcc ↑
aAcc ↑
mIoU ↑
mAcc ↑
SegFormer [ 51 ]
70.94
18.81
30.99
81.36
31.33
43.15
SegFormer [ 51 ] +AGLLDiff [ 35 ]
74.94
21.32
41.42
81.66
28.09
44.16
SegFormer [ 51 ] +Ours
78.90
27.31
45.58
85.31
35.13
50.75
Improve (%)
+5.28
+28.10
+10.04
+4.47
+25.06
+14.92
TABLE IX: Quantitative results of downstream semantic segmentation on the LOLv1 and LOLv2-real datasets. The segmentation predictions on the aligned normal-light images are used as reference masks for metric computation. Improve (%) denotes the relative improvement of SegFormer+Ours over SegFormer+AGLLDiff for each metric.
Low-Light Image Enhancement (LLIE) has long been a challenging problem in low-level vision, as insufficient illumination often leads to low contrast, detail loss, and noise. Recent studies show that deep learning-based Retinex theory can effectively decouple illumination and reflectance. However, existing methods frequently suffer from over-enhancement or color distortion, and often assume uniform noise or ideal lighting. To address these limitations, we propose InterLight, a novel framework that systematically excavates and operationalizes intrinsic illumination priors for LLIE.Our core insight is that robust enhancement requires not just estimating illumination, but constructing an illumination-aware pipeline. We first inject sensor-level illumination-response priors via physics-guided augmentation, then represent the degradation through adaptive prompts conditioned on the scene's latent illumination state. This explicit representation directly guides a luminance-gated intrinsic memory mechanism to selectively compensate for information loss, prioritizing reconstruction in dark regions while preserving fidelity in bright ones. Finally, the entire process is regularized by a self-supervised consistency objective that distills illumination-invariant features. By deeply exploiting intrinsic illumination priors, our method achieves clearer textures and more visually coherent enhancement results. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our approach. Code is available at: https://github.com/House-yuyu/InterLight.
Ziqi Wang, Xu Zhang, Laibin Chang +3
National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · Department of Computer Science, University of Macau · Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence +1
In this paper, we propose a zero-reference diffusion-based framework, named ZeroIDIR, for illumination degradation image restoration, which decouples the restoration process into adaptive illumination correction and diffusion-based reconstruction while being trained solely on low-quality degraded images. Specifically, we design an adaptive gamma correction module that performs spatially varying exposure correction to generate illumination-corrected only representations to mitigate exposure bias and serve as reliable inputs for subsequent diffusion processes, where a histogram-guided illumination correction loss is introduced to regularize the corrected illumination distribution toward that of natural scenes. Subsequently, the illumination-corrected image is treated as an intermediate noisy state for the proposed perturbed consistency diffusion model to reconstruct details and suppress noise. Moreover, a perturbed diffusion consistency loss is proposed to constrain the forward diffusion trajectory of the final restored image to remain consistent with the perturbed state, thus improving restoration fidelity and stability in the absence of supervision. Extensive experiments on publicly available benchmarks show that the proposed method outperforms state-of-the-art unsupervised competitors and is comparable to supervised methods while being more generalizable to various scenes. Code is available at https://github.com/JianghaiSCU/ZeroIDIR.
Hai Jiang, Zhen Liu, Yinjie Lei +3
School of Aeronautics and Astronautics, Sichuan University · University of Electronic Science and Technology of China · College of Electronics and Information Engineering, Sichuan University
Existing deep learning-based low-light enhancement methods are typically trained on limited datasets with single enhancement targets, which restricts their generalization ability and controllability in real-world applications. To overcome these limitations, we propose ControlLight, a controllable, consistent, and generalizable framework for low-light enhancement. We first construct a large-scale dataset of real-world degraded images with continuous illumination-strength supervision. To further ensure consistent outputs under different control strengths, we introduce a misalignment-aware weighted flow matching loss that preserves image structure across continuous enhancement strengths. ControlLight allows users to edit real-world degraded low-light images toward satisfactory enhancement results by flexibly controlling the strength while preserving visual consistency and realism. Extensive experiments show that ControlLight achieves state-of-the-art performance against existing low-light enhancement approaches while demonstrating strong continuous controllability and generalization to real-world scenarios.
Yufeng Yang, Jianzhuang Liu, Jisheng Chu +4
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Zhejiang University