Organizations: School of Automation, Hangzhou Dianzi University, Hangzhou 310018, China · College of Artificial Intelligence, Jiaxing University, Jiaxing 314001, China · School of Artificial Intelligence at Dalian University of Technology, Dalian 116000, China · Department of Computer Science at City University of Hong Kong, Hong Kong, China
Retinex theory provides a principled foundation for low-light image enhancement, inspiring numerous learning-based methods that integrate its principles. However, existing methods exhibits limitations in accurately decomposing reflectance and illumination components. To address this, we propose a Retinex-Guided Transformer~(RGT) model, which is a two-stage model consisting of decomposition and enhancement phases. First, we propose a latent space decomposition strategy to separate reflectance and illumination components. By incorporating the log transformation and 1-pixel offset, we convert the intrinsically multiplicative relationship into an additive formulation, enhancing decomposition stability and precision. Subsequently, we construct a U-shaped component refiner incorporating the proposed guidance fusion transformer block. The component refiner refines reflectance component to preserve texture details and optimize illumination distribution, effectively transforming low-light inputs to normal-light counterparts. Experimental evaluations across four benchmark datasets validate that our method achieves competitive performance in low-light enhancement and a more stable training process.
Figures & tables
Fig. 1: Comparison between the typical Retinex decomposition strategy and our method in deep learning frameworks.
Fig. 2: The overview of the RGT model. (a) Decomposition takes a paired low-light/normal-light image as input during training to decompose the image into R and L components. (b) RGT extracts R and L using the pre-trained decomposition network (frozen parameters) and enhances R and L separately via the Enhancer network (non-shared parameters). Then, reconstructing the final enhanced result by combining processed R and L .
Fig. 3: The architecture of the GFTB.
Fig. 4: Schematic illustration of illumination-reflectance component swapping between two content-identical images under different lighting conditions. Each input image is decomposed into reflectance ( R ) and illumination ( L ) components within the shallow spatial space. By exclusively exchanging the L components between paired images, we can synthesize four novel recomposed results. Ideally, swapping only the L components is sufficient to alter the brightness of the images.
Methods
LOLv2_real
LOLv2_syn
frec(Rl,Ll)
frec(Rl,Ln)
frec(Rn,Ll)
frec(Rn,Ln)
frec(Rl,Ll)
frec(Rl,Ln)
frec(Rn,Ll)
frec(Rn,Ln)
RetinexNet [ 4 ]
38.56
23.90
35.47
29.50
36.44
20.37
26.59
30.28
RetinexDIP [ 58 ]
16.61
14.91
14.49
18.05
23.75
14.43
23.52
20.47
URetinexNet [ 59 ]
34.85
19.40
31.47
24.76
26.78
17.39
25.36
22.89
RGT
69.45
54.77
68.34
59.05
58.22
51.11
56.82
53.72
TABLE I: The PSNR of comparative decomposition on LOLv2 dataset. frec represents the operation of reconstructing the image from R and L components. According to RetinexNet [ 4 ] requirements for the decomposed R and L components, the reconstructed images from R and L should match the input image corresponding to L .
Fig. 5: Visual comparisons of the decomposed results by different methods on LOLv2_real. The images within the green boxes should be same and the images within the red boxes should be same.
Methods
LOLv2_real
LOLv2_syn
HDR+(480p)
SDSD-indoor
SICE
Complexity
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
FLOPs(G)
Params(M)
Time(ms)
RetinexNet [ 4 ]
16.52
0.674
19.11
0.805
19.54
0.678
25.88
0.832
19.71
0.686
587.47
0.84
3.36
LLFlow [ 70 ]
20.41
0.829
23.14
0.930
22.96
0.762
22.10
0.801
19.99
0.685
358.40
17.42
122.70
MirNetv2 [ 71 ]
22.00
0.818
24.86
0.925
22.87
0.765
29.68
0.870
24.29
0.767
140.21
5.86
61.69
SNR-Net [ 72 ]
21.48
0.849
24.14
0.928
22.50
0.764
29.44
0.894
21.78
0.729
26.35
4.01
6.25
GSAD [ 73 ]
22.20
0.811
22.46
0.897
18.03
0.781
–
–
15.11
0.553
134.04
18.42
164.99
TABLE II: Quantitative comparisons PSNR/SSIM on LOLv2 [ 5 ] , HDR+ [ 6 ] , SDSD(indoor) [ 7 ] and SICE [ 8 ] datasets. The highest result is in bold while the second highest result is underlined .
Fig. 6: Visual comparisons of the enhanced results by different methods on LOLv2_real (top) and LOLv2_syn (bottom).
Fig. 7: Visual comparisons of the enhanced results by different methods on HDR+(480p) (top), SICE (middle) and SDSD-indoor (bottom)
Methods
LOLv2_real
LOLv2_syn
HDR+(480p)
SDSD-indoor
SICE
SNR-Net [ 72 ]
0.122
0.061
0.238
0.230
0.206
CSPN [ 75 ]
0.156
0.055
0.071
0.181
0.235
IGDFormer [ 77 ]
0.158
0.074
0.086
0.227
0.237
AttnFlow [ 76 ]
0.146
0.050
0.064
0.223
0.298
RetinexMamba [ 78 ]
0.147
0.062
0.073
0.184
0.230
CIDNet [ 79 ]
0.131
0.055
0.086
0.198
0.258
TABLE III: Quantitative comparisons LPIPS ↓ on LOLv2 [ 5 ] , HDR+(480p) [ 6 ] , SDSD(indoor) [ 7 ] and SICE [ 8 ] datasets. The highest result is in bold while the second highest result is underlined .
Methods
DICM
LIME
MEF
NPE
VV
CLIP-IQA ↑
NIQE ↓
CLIP-IQA ↑
NIQE ↓
CLIP-IQA ↑
NIQE ↓
CLIP-IQA ↑
NIQE ↓
CLIP-IQA ↑
NIQE ↓
SNR-Net [ 72 ]
0.599
4.71
0.665
5.74
0.633
4.18
0.544
4.32
0.304
9.87
CSPN [ 75 ]
0.659
4.10
0.626
4.12
0.604
4.42
0.668
3.97
0.297
3.24
IGDFormer [ 77 ]
0.652
3.77
0.665
3.96
0.688
3.51
0.530
4.30
0.313
2.99
AttnFlow [ 76 ]
0.730
4.54
0.712
4.46
0.746
4.64
0.683
4.24
0.352
3.00
RetinexMamba [ 78 ]
0.665
3.79
0.691
3.96
0.673
3.35
0.573
3.92
0.384
2.90
TABLE IV: Quantitative comparisons CLIP-IQA ↑ /NIQE ↓ on DICM [ 60 ] , LIME [ 61 ] , MEF [ 62 ] , NPE [ 63 ] and VV [ 64 ] datasets. The highest result is in bold while the second highest result is underlined .
Fig. 8: Visual comparisons of the enhanced results by different methods on unpaired dataset (DICM, LIME, MEF, NPE and VV).
Fig. 9: Visual comparison of various methods on real-world images from various mobile devices.
Methods
frec(Rl,Ll)
frec(Rl,Ln)
frec(Rn,Ll)
frec(Rn,Ln)
PSNR
SSIM
(0)
40.35
30.84
39.91
38.35
21.76
0.844
(1)
50.16
23.56
42.76
38.52
22.30
0.844
(2)
52.73
48.19
51.77
51.58
23.27
0.811
(3)
51.12
22.83
39.03
50.53
22.63
0.832
Ours
69.45
54.77
68.34
59.05
23.80
0.896
TABLE V: Ablation study on decomposition strategies using the LOLv2_real benchmark. Subscripts l/n denote components derived from low-light and normal images respectively.
Fig. 10: The overview of four variant decomposition strategies.
Inputs
Methods
frec(Rl,Ll)
frec(Rl,Ln)
frec(Rn,Ll)
frec(Rn,Ln)
Avg
Low ↔ Middle
(1)
57.27
46.19
46.18
57.33
51.74
(2)
64.74
59.52
62.16
61.03
61.86
Ours
68.00
61.92
66.86
62.73
64.88
Middle ↔ High
(1)
58.38
41.27
39.42
50.39
47.37
(2)
61.03
57.39
58.29
59.18
58.97
Ours
62.73
57.62
61.38
58.78
60.13
TABLE VI: Ablation study on decomposition strategies using the Kalantri dataset. Based on the types of input image pairs, the experiments can be divided into three groups:(i) Low ↔ Middle; (ii) Middle ↔ High; (iii) Low ↔ High. Each group employs three decomposition strategies for the experiments, including the classical Retinex strategy (variant (1)), variant (2), and the full strategy (Ours).
Methods
frec(Rl,Ll)
frec(Rl,Ln)
frec(Rn,Ll)
frec(Rn,Ln)
Avg
RetinexNet
38.56
23.90
35.47
29.50
31.86
RetinexNet*
58.27
45.09
57.84
50.23
52.86
URetinexNet
34.85
19.40
31.47
24.76
27.62
URetinexNet*
56.75
45.35
56.35
47.13
51.40
Ours
69.45
54.77
68.34
59.05
62.91
TABLE VII: Ablation study on decomposition strategies using the LOLv2_real benchmark. Subscripts l/n denote components derived from low-light and normal images respectively. * denotes the model employs the latent space decomposition strategy.
Fig. 11: Comparison of training stability between additive and multiplicative decompositions. Each model is trained 10 times, and the figure shows the mean and variance of the loss at each epoch.
Fig. 12: The overview of three GFTB variants and complete GFTB.
Different GFTB structures
Metrics
Variants
R
L
(a) SA
(b) ⊙V
(c) ⊙fin
(d) CA
Smean
Smax
Smean
Smax
PSNR
SSIM
✓
22.99
0.837
✓
✓
✓
23.03
0.792
✓
✓
✓
21.85
0.815
✓
✓
✓
23.53
0.855
TABLE VIII: Ablation study on different GFTB structures.
This paper proposes a neural network for low light image enhancement (LLIE) based on retinex theory to make LLIE robust for various dynamic range scenes. The retinex theory is an image formulation model inspired by a human color perception hypothesis, where a low light image is decomposed into intrinsic color context (i.e., reflectance map) and scene-dependent illumination (i.e., illumination map). Due to non-uniqueness of its decomposition, existing retinex-based LLIE methods often fail to achieve stable decomposition, which lead to over-enhancement. Typically, they are sensitive to the dynamic ranges that vary in different lighting conditions. To tackle this issue, we propose WREN: An LLIE neural network with double U-Net-like structures. WREN consists of two U-Net-like sub-networks. The first network has one encoder and two decoders that decompose an input image into the reflectance and illumination maps. The second network with a customized Transformer block between an encoder and a decoder only enhances the illumination map obtained from the first network: This completely follows the assumption of the retinex theory. Finally, the enhanced illumination map is recombined with the reflectance map. The network is trained end-to-end with a scale-invariant loss function, which gives robustness against the illumination scaling. Numerical results show that our method achieves the state-of-the-art performance across multiple datasets. Our code is available online.
Reina Kaneko, Junya Hara, Hiroshi Higashi +1
The University of Osaka, Osaka, Japan · Kansai University, Osaka, Japan
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
Low-light image enhancement is challenging due to complex degradations, including amplified noise, artifacts, and color distortion. While Retinex-based deep learning methods have achieved promising results, they primarily rely on single-modality RGB information. We propose M2Retinexformer (Multi-Modal Retinexformer), a novel framework that extends Retinexformer by incorporating depth cues, luminance priors, and semantic features within a progressive refinement pipeline. Depth provides geometric context that is invariant to lighting variations, while luminance and semantic features offer explicit guidance on brightness distribution and scene understanding. Modalities are extracted at multiple scales and fused through cross-attention, with adaptive gating dynamically balancing illumination-guided self-attention and cross-attention based on the reliability of auxiliary cues. Evaluations on the LOL, SID, SMID, and SDSD benchmarks demonstrate overall improvements over Retinexformer and recent state-of-the-art methods. Code and pretrained weights are available at https://github.com/YoussefAboelwafa/M2Retinexformer
Youssef Aboelwafa, Hicham G. Elmongui, Marwan Torki