Organizations: College of Computer Science and Technology, Jilin University, Changchun, Jilin, 130012, China, and the Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, Jilin 130012, China. · College of Software, Jilin University, Changchun, Jilin, 130012, China. · School of Computer and Artificial Intelligence, XiangnanUniversity, Chenzhou, 423300, China. · College of Communication Engineering, Jilin University, Changchun, Jilin, 130012, China.
Frequency-domain methods have shown strong potential for low-light image enhancement (LLIE) because they can effectively regulate global illumination. However, most existing methods assume that amplitude and phase can be restored independently, while overlooking their inherent spectral coupling. This artificial decoupling often leads to information inconsistency and structural distortions. To address this issue, we propose a frequency-spatial framework that jointly models spectral interaction and reliable prior guidance. We first empirically analyze the coupling between amplitude and phase and develop an Amplitude-Phase Adaptive Compensation Fusion (APACF) module. Instead of enforcing a hard separation between the two components, APACF enables their bidirectional interaction, resulting in spectrally coherent restoration. In addition, directly extracting priors from raw low-light inputs is unreliable, since severe noise and underexposure destroy critical visual content. To alleviate this problem, we introduce an Endogenous Reliability Prior (ERP). Specifically, we treat an intermediate result, preliminarily restored in the frequency domain and already close to the ground-truth illumination, as a high-fidelity source for extracting complementary pseudo-infrared and gradient cues. Based on ERP, we further design a Dual-Modal Guided Refinement Network (DGRN) to inject these purified structural cues into the decoder, thereby recovering missing fine textures more accurately. Extensive experiments on benchmark datasets show that our method achieves state-of-the-art performance and consistently surpasses existing methods in both quantitative metrics and visual quality. Our code is available at https://github.com/bywlzts/Amplitude-Phase-Coupling-Net.
Figures & tables
Fig. 1: Reprojection consistency test. Combining the clean amplitude Agt with the degraded phase Plow produces a mixed image Imixed with severe artifacts. Reprojecting Imixed to the frequency domain yields clear discrepancies between the original and reconstructed components, e.g., Agt=A^gt , revealing strong coupling between amplitude and phase and challenging the assumption that they can be restored independently.
Fig. 2: Validation of ERP. Left: t -SNE visualization of feature distributions for the low-light input (red), the Stage I output (blue), and the ground truth (green). Right: Comparison of Pseudo-Infrared (PI) features extracted from the low-light input and the Stage I output.
Fig. 3: Overview of the proposed framework. Stage I: APACF performs coupled amplitude-phase restoration to produce Is1 , from which ERP is extracted. Stage II: DGRN refines details under ERP guidance to generate the enhanced output Iout .
Fig. 4: Gradient-Guided Enhancement Block (GGEB). MDAM first aligns PI and gradient features across channel, spatial, and pixel dimensions. SGCA then performs gradient-guided cross-attention to produce the endogenous prior feature Fen .
Methods
LOL-v1
LOL-v2-Syn
LSRW-Huawei
LSRW-Nikon
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
MIRNet [ 35 ]
24.14
0.8305
0.2502
22.52
0.8997
0.0568
19.98
0.6085
0.2154
17.10
0.5022
0.2170
Kind++ [ 36 ]
18.97
0.8042
0.1756
21.17
0.8814
0.0678
15.43
0.5695
0.2366
14.79
0.4749
0.2111
FourLLIE [ 12 ]
20.99
0.8071
0.0952
24.65
0.9192
0.0389
21.11
0.6256
0.1825
17.82
0.5036
0.2150
UHDFour [ 13 ]
22.89
0.8147
0.0934
23.64
0.8998
0.0341
19.39
0.6006
0.2466
17.94
0.5195
0.2546
DMFourLLIE [ 14 ]
22.98
0.8273
0.0792
25.74
0.9308
0.0251
21.09
0.6328
0.1804
17.04
0.5274
0.2178
TABLE I: Quantitative comparison on LOL-v1 [ 33 ] , LOL-v2-Syn [ 33 ] , LSRW-Huawei [ 34 ] and LSRW-Nikon [ 34 ] . Bold and underline indicate the best and second-best results, respectively.
Fig. 5: Visual comparison with state-of-the-art methods on the LOL-v1 dataset.
Fig. 6: Visual comparison with state-of-the-art methods on the LOL-v2-Syn dataset.
Fig. 7: Visual comparison with state-of-the-art methods on the LSRW-Huawei dataset.
Fig. 8: Visual comparison with state-of-the-art methods on the LSRW-Nikon dataset.
Fig. 9: Visual comparison on MEF and NPE datasets.
Methods
LIME
VV
DICM
NPE
MEF
AVG
GLARE [ 45 ]
4.343
4.262
4.858
5.467
3.906
4.567
LYTNet [ 46 ]
4.540
4.639
4.187
4.372
3.959
4.339
LANet [ 47 ]
3.680
4.670
3.902
4.856
3.931
4.208
FourLLIE [ 12 ]
3.814
4.222
4.365
4.914
3.831
4.229
Wave-Mamba [ 37 ]
4.451
4.713
4.568
4.545
4.767
4.608
Retinexformer [ 7 ]
3.442
3.719
4.013
3.891
3.730
3.759
TABLE II: Quantitative comparison using NIQE metric (lower is better) on unpaired real-world datasets.
Fig. 10: Comparison of computational efficiency in terms of FLOPs and average inference time. Our method achieves the lowest inference time while maintaining a moderate computational cost.
Model
Stage-I
Stage-II
PSNR ↑
SSIM ↑
LPIPS ↓
A
✓
×
17.52
0.7164
0.1496
B
×
✓
25.62
0.9330
0.0240
Ours
✓
✓
26.42
0.9371
0.0222
TABLE III: Ablation study on the effectiveness of the two-stage architecture.
Configuration
PSNR ↑
SSIM ↑
LPIPS ↓
w/o Frequency Branch
25.62
0.9330
0.0240
w/ Hard Splitting
24.92
0.9311
0.0270
w/ APACF (Ours)
26.42
0.9371
0.0222
TABLE IV: Ablation study of APACF.
Prior Source
PSNR ↑
SSIM ↑
LPIPS ↓
Raw-Prior (Input)
23.15
0.9139
0.0418
GT-Prior (Reference)
19.10
0.8664
0.0600
Depth Map [ 61 ]
24.87
0.9294
0.0342
Edge Map [ 62 ]
23.96
0.9253
0.0384
ERP (Ours)
26.42
0.9371
0.0222
TABLE V: Ablation study on different prior sources.
Model Variant
GGEB
PIGE
PSNR ↑
SSIM ↑
LPIPS ↓
Baseline
×
×
25.51
0.9345
0.0251
w/ GGEB Only
✓
×
25.44
0.9297
0.0260
w/ PIGE Only
×
✓
25.31
0.9352
0.0244
Ours
✓
✓
26.42
0.9371
0.0222
TABLE VI: Ablation study of GGEB and PIGE.
Fig. 11: Visual ablation comparison.
Fig. 12: Visualization of intermediate features under different prior settings, including no prior (RAW), infrared-only prior, gradient-only prior, and the proposed ERP.
Fig. 13: Statistical evidence of bidirectional spectral coupling. (a) Replacing only the low-light phase with a normal-light one shifts the pixel intensity distribution toward brighter regions, indicating that phase also affects illumination. (b) Combining the ground-truth amplitude (Ahigh) with the low-light phase still yields a gradient distribution that deviates from the GT.
Fig. 14: Visual ablation comparison for APACF.
Fig. 15: Visual ablation comparison of different prior sources.
Fig. 16: Qualitative comparison of semantic segmentation results on the NightCity dataset. Red boxes highlight representative local regions.
Fig. 17: Qualitative comparison of object detection results in low-light scenes on the ExDark dataset.
Fig. 18: Qualitative comparison of face detection results on the DarkFace dataset.
Fig. 19: Visual limitation analysis on real-world scenes.
Low-light image enhancement (LLIE) aims to improve image quality and clarity in diverse and demanding low-illumination environments. However, existing deep learning-based LLIE methods struggle to accurately capture real-world illumination and restore texture details, largely because their algorithmic strengths remain underutilized. To address these issues, we present a supervised frequency domain deep learning network for LLIE, named multi-scale attention combined with the Fourier transform (MSFT) which adopts a U-shaped, one-stage architecture that infuses guidance from low-light images into the network by channeling it through multi-scale attention. We further fuse the amplitude information from priori channels with that of the low-light image in MSFT's self-created module, and carry out multi-scale guidance along with the network. Subsequently, to better enhance the faint feature, such as fine content and textures, and to better fuse global context confidence in the decoding stage, we separately introduce a multi-shape synergistic attention and a lightweight network that effectively integrate information in high-dimensional space to embed into the superlative feature space channel containing rich texture information. Extensive experiments conducted on LOL, SID, SMID, and SDSD datasets demonstrate that MSFT significantly outperforms state-of-the-art competitors. For example, compared with Retinexformer, our method achieves a peak signal-to-noise ratio of up to 41.76 decibels on the SDSD-outdoor dataset with an increase of 11.92 decibels and a structural similarity index of 0.988 with a 13.80% improvement.
Wenbin Du, Jian Long, Zhu Cao
aKey Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, 200237, China · bCollege of Electronics and Information Engineering, Tongji University, Shanghai 201804, China · cShanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai 201210, China
Existing Ultra-High-Definition (UHD) low-light image enhancement (LLIE) methods typically rely on either spatial- or frequency-domain modeling, or independently process different frequency components for efficient restoration. However, they often fail to jointly exploit complementary spatial--frequency information and the inherent relationship between illumination and structural information across frequencies, leading to inconsistent illumination, residual noise, and detail artifacts. In this paper, we propose a novel Bidirectional Cross-frequency Domain Interaction Network (BiFreqNet) for UHD LLIE. Specifically, BiFreqNet incorporates a Dual-domain Feature Interaction (DFI) module into the multi-scale representations of a pretrained variational autoencoder (VAE) to perform reciprocal enhancement between low- and high-frequency features. DFI consists of a Spatial and Spectral-Aware Low-Frequency Enhancement (S2LFE) block and a Cross-Frequency Enhancement (CFE) block. S2LFE exploits high-frequency structural cues to guide low-frequency restoration and jointly models spatial and spectral information to recover global exposure and spatially varying illumination while suppressing sensor noise. CFE further leverages the enhanced low-frequency contextual information to guide adaptive high-frequency refinement, enabling faithful reconstruction of structures and fine details. Extensive experiments on four general and UHD LLIE benchmarks demonstrate that BiFreqNet achieves competitive or superior performance compared with state-of-the-art methods.
Bo Fu, Hai Chi, Liyan Wang +5
School of Computer and Artificial Intelligence, Liaoning Normal University, Dalian 116081, China · School of Mathematical Sciences, Dalian University of Technology, Dalian 116024, China · School of Computer Sciences, Dalian University of Technology, Dalian 116024, China +3
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