Anchor then Polish for Low-light Enhancement

Tianle Du; Mingjia Li; Hainuo Wang; Xiaojie Guo

cs.CVMar 16, 2026

Anchor then Polish for Low-light Enhancement

Authors: Tianle Du, Mingjia Li, Hainuo Wang, Xiaojie Guo

Organizations: School of Software, Tianjin University, Tianjin, China

Abstract

Low-light image enhancement is challenging due to entangled degradations, mainly including poor illumination, color shifts, and texture interference. Existing methods often rely on complex architectures to address these issues jointly but may overfit simple physical constraints, leading to global distortions. This work proposes a novel anchor-then-polish (ATP) framework to fundamentally decouple global energy alignment from local detail refinement. First, macro anchoring is customized to (greatly) stabilize luminance distribution and correct color by learning a scene-adaptive projection matrix with merely 12 degrees of freedom, revealing that a simple linear operator can effectively align global energy. The macro anchoring then reduces the task to micro polishing, which further refines details in the wavelet domain and chrominance space under matrix guidance. A constrained luminance update strategy is designed to ensure global consistency while directing the network to concentrate on fine-grained polishing. Extensive experiments on multiple benchmarks show that our method achieves state-of-the-art performance, producing visually natural and quantitatively superior low-light enhancements.

Figures & tables

Figure 1: (a) Task-level energy decomposition of low-light degradations. (b) Residual distribution shift after global energy anchoring (GEA), illustrating error compression and re-centering. (c) PSNR comparison across five benchmarks. (d) Visual validation of different global energy alignment manners, where the Ideal GEA uses the optimal matrix specific to the given image.

Figure 2: (a) Architecture of the proposed ATP framework. (b) Global Energy Anchoring Module (GEAM). (c) Detail Polishing Module (DPM).

Method	FPS $\uparrow$ (600 $\times$ 400)	Params $\downarrow$ (M)	LOLv1			LOLv2-Real			LOLv2-Synthetic
Method	FPS $\uparrow$ (600 $\times$ 400)	Params $\downarrow$ (M)	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
KinD [ 58 ]	2.75	8.02	19.427	0.7135	0.2391	17.544	0.6690	0.3750	17.137	0.7620	0.2550
ZeroDCE [ 7 ]	629.79	0.075	14.861	0.5588	0.3351	18.059	0.5736	0.3125	17.756	0.8163	0.1681
MIRNet [ 53 ]	6.50	31.79	20.566	0.7722	0.2542	21.281	0.7920	0.3538	21.741	0.8780	0.1383
RUAS [ 22 ]	159.45	0.003	16.405	0.4996	0.2701	15.326	0.4876	0.3097	13.402	0.6444	0.3640
SNR-Aware [ 46 ]	89.74	4.01	23.005	0.8240	0.1640	21.103	0.8390	0.1690	24.173	0.9240	0.0640

Table 1: Quantitative comparison on LOLv1 [ 43 ] and LOLv2 [ 50 ] datasets. Best results are in bold and second best are underlined . All results are reported using the checkpoint that achieves the highest PSNR without using the GT mean strategy.

Panel of Figure 3 — Figure 3: Visual comparisons of the enhanced results by different methods on LOLv1 [ 43 ] , LOLv2-Real [ 50 ] and LOLv2-Synthetic [ 50 ] . Zoom in for the best view.

Method	FPS $\uparrow$ (600 $\times$ 400)	Params $\downarrow$ (M)	LSRW-Nikon			LSRW-Nikon-Reg			UHD-LL
Method	FPS $\uparrow$ (600 $\times$ 400)	Params $\downarrow$ (M)	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
ZeroDCE [ 7 ]	629.79	0.075	15.028	0.4183	0.2438	17.236	0.6966	0.2104	17.071	0.6391	0.5136
MIRNet [ 53 ]	6.50	31.79	17.377	0.5015	0.5388	18.966	0.6105	0.5271	20.550	0.8067	0.4351
UFormer [ 40 ]	8.12	50.88	17.086	0.5003	0.4627	18.788	0.6165	0.4680	19.306	0.8053	0.4360
Retformer [ 3 ]	28.45	1.53	17.303	0.5063	0.4897	19.716	0.6201	0.4734	22.217	0.8350	0.2815
CIDNet [ 49 ]	25.20	1.88	17.148	0.5038	0.2984	19.444	0.6599	0.3062	23.799	0.8746	0.2094

Table 2: Quantitative comparison on LSRW-Nikon [ 10 ] and UHD-LL [ 20 ] datasets. Best results are in bold and second best are underlined .

Panel of Figure 4 — Figure 4: Visual comparisons of the enhanced results by different methods on LSRW-Nikon [ 10 ] and UHD-LL [ 20 ] . Zoom in for the best view.

Panel of Figure 5 — Figure 5: Visual comparison on the five unpaired datasets. Zoom in for the best view.

Figure 6: Analysis of global energy alignment and CLU. (a) Reconstructed Y-channel DWT-LL components. (b) Absolute error maps. (c) PSNR and SSIM training curves with and without CLU, given the low global luminance error evidenced in (a) and (b).

Method	DICM		LIME		MEF		NPE		VV
Method	IQA $\uparrow$	NIQE $\downarrow$	IQA $\uparrow$	NIQE $\downarrow$	IQA $\uparrow$	NIQE $\downarrow$	IQA $\uparrow$	NIQE $\downarrow$	IQA $\uparrow$	NIQE $\downarrow$
RUAS [ 22 ]	2.04	7.05	2.11	5.37	2.69	5.41	1.63	7.09	2.27	4.75
MIRNet [ 53 ]	2.69	5.05	2.40	5.55	2.17	5.50	2.21	5.53	2.58	4.47
UFormer [ 40 ]	2.54	5.32	2.49	5.32	2.31	5.52	2.43	5.96	2.57	5.39
Retinexformer [ 3 ]	3.18	4.13	3.25	4.21	3.05	4.18	2.95	4.21	2.92	3.26
GSAD [ 11 ]	3.25	4.15	3.31	4.60	3.40	4.45	3.36	4.72	3.01	4.10

Table 3: No-reference evaluation on five low-light datasets. Best results are in bold and second best are underlined .

Set	Components			MG	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Set	GEAM	DPM	CLU	MG	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
A	$\times$	$\surd$	$\times$	-	22.309	0.8523	0.1268
B	$\times$	$\surd$	$\surd$	-	15.069	0.6652	0.3322
C	$\surd$	$\surd$	$\times$	$\surd$	22.718	0.8567	0.1304
D	$\surd$	$\times$	$\times$	$\times$	24.001	0.7279	0.2490
E	$\surd$	$\surd$	$\surd$	$\times$	24.839	0.8811	0.1054

Table 4: Ablation study of different components.

Set	Matrix shape (DoF)	Formulation	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
A	$1\times 1~(1)$	$a\cdot\mathbf{I}$	23.451	0.8642	0.1149
B	Diag $(\mathbf{d})$ + $\mathbf{b}\,(6)$	$\mathbf{A}\cdot\mathbf{I}+\mathbf{b}$	23.014	0.8764	0.1063
C	$3\times 3~(9)$	$\mathbf{A}\cdot\mathbf{I}$	21.674	0.8651	0.1265
D	$3\times 4~(12)$	$\mathbf{A}\cdot\mathbf{I}+\mathbf{b}$	25.140	0.8851	0.0959

Table 5: Ablation study on Matrix shape (Degrees of Freedom, DoF).

DWT applied on	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
YUV	24.318	0.8768	0.1028
RGB	24.047	0.8777	0.1049
Y	25.140	0.8851	0.0959

Table 6: Ablation of applying DWT on different channels.

Loss	GEAM			DPM					PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Loss	$\lambda_{\text{rec}}$	$\lambda_{\text{mat}}$	$\lambda_{\text{diag}}$	$\lambda_{\text{color}}$	$\lambda_{\text{freq}}$	$\lambda_{\text{lum}}$	$\lambda_{\text{ssim}}$	$\lambda_{\text{vgg}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
A	1	1	0.01	1	1	0.01	0.5	0.01	24.912	0.8795	0.1021
B	1	1	0.1	1	1	0.1	0.5	0.01	24.433	0.8802	0.1090
C	1	1	0.1	1	1	0.01	0.3	0.01	24.355	0.8818	0.0971
D	1	1	0.1	1	1.5	0.01	0.5	0.01	24.674	0.8827	0.1002
E	1	0.5	0.2	1	1	0.01	0.5	0.01	23.651	0.8800	0.0978

Table 7: Ablation study on loss weights.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Observations after applying GEA to more datasets.

Figure 2: Matrix structure of GEAM prediction on LOLv2-Real. Where the $\rho$ indicates a monotonic relationship between matrix elements and input luminance. The closer $\left|\rho\right|$ is to 1, the stronger the relationship.

Figure 3: Parameter Embedding Module. Red indicates that the feature is amplified, black indicates that the feature is suppressed, and blue indicates that the feature remains unchanged.

Figure 4: Visual analysis of the proposed Matrix Guidance mechanism.

Panel of Figure 6 — Figure 6: More visual comparisons of the enhanced results by different methods on LOLv2-Real [ 50 ] . Zoom in for the best view.

Panel of Figure 7 — Figure 7: More visual comparisons of the enhanced results by different methods on LOLv2-Synthetic [ 50 ] . Zoom in for the best view.

Panel of Figure 8 — Figure 8: More visual comparisons of the enhanced results by different methods on LSRW-Nikon [ 10 ] and UHD-LL [ 20 ] . Zoom in for the best view.

Panel of Figure 9 — Figure 9: Qualitative comparison in challenging real-world low-light scenarios. In scenes characterized by significant spatial variations in illumination and extreme contrast, ATP and existing methods both encounter difficulties in perfect color restoration. However, ATP demonstrates superior robustness in preserving structural details and suppressing artifacts.

Explore similar work

ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement
May 25, 2026Yufeng Yang, Jianzhuang Liu, Jisheng Chu +4Image Enhancement Low-Light Image Enhancement
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Zhejiang University
Enlight: Fast Low-Light Image Enhancement via Multi-Objective Optimization and Shadow-Aware Refinement
Jun 19, 2026Nirjhor Datta, M. Sohel RahmanImage Enhancement Inference-Time Optimization
Department of Computer Science and Engineering, BRAC University, Dhaka, 1205, Bangladesh · Department of Computer Science and Engineering, Bangladesh University of Engineering and Technology, Dhaka, 1205, Bangladesh
InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement
May 19, 2026Ziqi Wang, Xu Zhang, Laibin Chang +3Image Enhancement Low-Light Image Enhancement
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

May 25, 2026J/K move · Enter open · S save

ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement

Yufeng Yang, Jianzhuang Liu, Jisheng Chu, Yuqi Peng+3

Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Zhejiang University

Image Enhancement Low-Light Image Enhancement

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement

Enlight: Fast Low-Light Image Enhancement via Multi-Objective Optimization and Shadow-Aware Refinement

InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement