VCR: Variance-Driven Channel Recalibration for Robust Low-Light Enhancement

Zhixin Cheng; Fangwen Zhang; Xiaotian Yin; Baoqun Yin; Haodian Wang

cs.CVMar 10, 2026

VCR: Variance-Driven Channel Recalibration for Robust Low-Light Enhancement

Authors: Zhixin Cheng, Fangwen Zhang, Xiaotian Yin, Baoqun Yin, Haodian Wang

Organizations: School of Information Science and Technology, University of Science and Technology of China, Hefei 230027, China · Institute of Advanced Technology, University of Science and Technology of China, Hefei 230027, China · CHN Energy Digital Intelligence Technology Development (Beijing) Co., LTD., Beijing, China

Abstract

Most sRGB-based LLIE methods suffer from entangled luminance and color, while the HSV color space offers insufficient decoupling at the cost of introducing significant red and black noise artifacts. Recently, the HVI color space has been proposed to address these limitations by enhancing color fidelity through chrominance polarization and intensity compression. However, existing methods could suffer from channel-level inconsistency between luminance and chrominance, and misaligned color distribution may lead to unnatural enhancement results. To address these challenges, we propose the Variance-Driven Channel Recalibration for Robust Low-Light Enhancement (VCR), a novel framework for low-light image enhancement. VCR consists of two main components, including the Channel Adaptive Adjustment (CAA) module, which employs variance-guided feature filtering to enhance the model's focus on regions with high intensity and color distribution. And the Color Distribution Alignment (CDA) module, which enforces distribution alignment in the color feature space. These designs enhance perceptual quality under low-light conditions. Experimental results on several benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance compared with existing methods.

Figures & tables

Fig. 1: (a) Different feature channels focus on different regions. By selectively filtering channels, regions with consistent brightness and color distributions are enhanced, leading to improved overall enhancement performance. (b) Consistency in color space distribution helps the image achieve a more natural appearance, leading to more realistic and visually pleasing enhancement results.

Fig. 2: The overall pipeline of the VCR process begins by transforming the input into the HVI space. Next, it is processed by the Channel Adaptive Adjustment module, which includes Variance-aware Channel Filtering and Triplet Channel Enhancement stage. These techniques aim to emphasize regions with a high consistency of luminance and chromaticity by filtering and enhancing the channels. After this recalibration, the features are refined, and the HV components are aligned with ground-truth statistics via the Color Distribution Alignment (CDA) module to mitigate color shifts. Finally, the enhanced output is reconstructed in the sRGB color space.

Fig. 3: Visual results of various methods on the LOL dataset. Regions highlighted by green and yellow boxes indicate differences in local details.

Fig. 4: Qualitative comparison of enhancement results on the unpaired dataset in difficult condition, generated by various methods.

Fig. 5: Qualitative comparison of enhancement results on the unpaired dataset in one difficult condition.

Fig. 6: (a) Qualitative comparison of TCE module. (b) Ablation visualization results of different space. (c)Visual comparison between prior methods and ours. The top row presents the RGB images, while the bottom row shows the corresponding discrete color histograms.

Fig. 7: (a) shows ablation results for varying quantities of CAA modules, while (b) presents results from experiments using different module types on the LOLv2-REAL dataset.

Fig. 8: Qualitative comparison of VCF module.

Fig. 9: More visual results of various methods on the LOLv1 dataset.

Methods	Complexity		LOLv1			LOLv2-Real			LOLv2-Synthetic
Methods	Params/M	FLOPs/G	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
RetinexNet [ 22 ]	0.84	584.47	18.915	0.427	0.470	16.097	0.401	0.543	17.137	0.762	0.255
KinD [ 43 ]	8.02	34.99	23.018	0.843	0.156	17.544	0.669	0.375	18.320	0.796	0.252
ZeroDCE [ 44 ]	0.075	4.83	21.880	0.640	0.335	16.059	0.580	0.313	17.712	0.815	0.169
RUAS [ 45 ]	0.003	0.83	18.654	0.518	0.270	15.326	0.488	0.176	13.765	0.638	0.305
LLFlow [ 46 ]	17.42	358.4	24.998	0.871	0.117	17.433	0.831	0.315	24.870	0.919	0.067

TABLE I: Quantitative results of PSNR

\uparrow

/SSIM

\uparrow

and LPIPS

\downarrow

on the LOL (v1 and v2) datasets. The best performance is in red , and the second best is in blue .

Fig. 10: More visual results of various methods on the LOLv2 and SICE dataset.

Methods	SICE		SID		Unpaired
	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	BRISQUE $\downarrow$	NIQE $\downarrow$
RetinexNet	12.424	0.613	15.695	0.395	23.286	4.558
ZeroDCE	12.452	0.639	14.087	0.090	26.343	4.763
RUAS	8.656	0.494	12.622	0.081	26.372	4.800
LLFlow	12.737	0.617	16.226	0.367	26.087	4.221
CIDNet	13.435	0.642	22.904	0.676	23.521	3.523

TABLE II: Quantitative result on SID, SICE and the five unpaired datasets (DICM, LIME, MEF, NPE, and VV). The top-ranking score is in red .

Exp.	CAA	TCE	$\mathbf{VCF}_{(\lambda_{\text{VCF})}}$	$\mathbf{CDA}_{(\lambda_{\text{CDA})}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
1					23.972	0.817	0.135
2	✓				24.158	0.825	0.129
3	✓	✓			24.364	0.839	0.116
4	✓	✓	✓		24.519	0.873	0.112
5	✓	✓	✓	✓	24.758	0.893	0.105

TABLE III: Ablation studies of modules.

Method	PSNR $\uparrow$	SSIM $\uparrow$
w/o TCE	24.683	0.874
TCE-Vanilla Conv	24.702	0.889
Ours	24.758	0.893

TABLE IV: Ablation of TCE module.

Methods	PSNR $\uparrow$	SSIM $\uparrow$
HVI (CIDNet)	24.111	0.871
HVI (Ours)	24.758	0.893
HSV (CIDNet)	21.349	0.801
HSV (Ours)	22.104	0.853

TABLE V: Comparison between CIDNet and our method in different color spaces on LOLv2-Real.

Method	PSNR $\uparrow$	SSIM $\uparrow$
w/o $F_{t1}$	24.613	0.881
w/o $F_{t2}$	24.562	0.879
w/o $F_{t3}$	24.598	0.877
Ours	24.758	0.893

TABLE VI: Ablation of three-branch structure.

Metrics	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Masking Ratio
Ratio=1/5	23.997	0.853	0.141
Ratio=1/4	24.516	0.872	0.126
Ratio=1/3	24.758	0.893	0.105
Ratio=1/2	22.963	0.804	0.207

TABLE VII: Ablation on the masking ratio of Variance‐aware Channel Filtering (VCF) module.

Setting	Similarity
CIDNet (HVI features vs. GT)	0.8788
C1 (small variance channels vs. GT)	0.8965
C2 (large variance channels vs. GT)	0.8549
Ours (final features vs. GT)	0.9009

TABLE VIII: Similarity comparison under different settings.

Fig. 11: (a) Failure cases of the proposed method under extreme nighttime scenarios. (b) Comparison visualization of SID dataset.

Fig. 12: More visual results of various methods on the unpaired dataset.

Explore similar work

Towards Color-Faithful Low-Light Image Enhancement via Adaptive Color Debiasing and Saturation Rectification
Aug 11, 2026Zhichen Yang, Rui Xu, Yuzhen Niu +3Image Enhancement Low-Light Image Enhancement
Fuzhou University, Fuzhou, Fujian, China
BC-IHV: Conditioning the Color Space for Stable Rectified-Flow Low-Light Enhancement
Aug 22, 2026Yi Ai, Zheng Chen, Yuanhao Cai +2Image Enhancement Low-Light Image Enhancement
Shanghai Jiao Tong University · Johns Hopkins University
Retrieval-Augmented Generation-Based Color Restoration for Low-Light Image Enhancement
Aug 8, 2026Li-Wei Lu, Shaou-Gang MiaouLow-Light Image Enhancement Image Restoration
Department of Electronic Engineering, Chung Yuan Christian University, Taoyuan 320, Taiwan

Aug 11, 2026J/K move · Enter open · S save

Towards Color-Faithful Low-Light Image Enhancement via Adaptive Color Debiasing and Saturation Rectification

Zhichen Yang, Rui Xu, Yuzhen Niu, Fusheng Li+2

Fuzhou University, Fuzhou, Fujian, China

Image Enhancement Low-Light Image Enhancement

Open paper AskPDF arXiv

Low-light imaging often introduces color bias caused by the low signal-to-noise ratio and the image formation process. Although recent low-light image enhancement methods have achieved strong brightness recovery, faithful color restoration remains challenging, manifesting as overall color bias together with local under- and over-saturation. To address this issue, we propose CAGE, a cylindrical color correction framework with adaptive color debiasing and gamut-harmonized saturation rectification for color-faithful low-light image enhancement. We first introduce AdaLAB, a cylindrical adaptive LAB color space that provides a decoupled and image-specific basis for uniform color correction. Building on this color space, we further develop AdaCCT, an adaptive cylindrical color transform with forward and inverse transforms for the conversion between RGB and AdaLAB color space, as well as necessary color debiasing and saturation rectification. The forward transform suppresses embedded color bias before backbone enhancement by reorganizing the chromatic distribution through chromatic-plane shifting and scaling, while the inverse transform achieves faithful saturation rectification through out-of-gamut lightness compensation. Extensive experiments on multiple benchmarks show that CAGE achieves more faithful color restoration, specifically reduces color bias and saturation abnormality, and delivers better overall visual quality across different low-light enhancement backbones. The code is available at https://yangzhichen763.github.io/CAGE/.

Abstract

Explore similar work

Towards Color-Faithful Low-Light Image Enhancement via Adaptive Color Debiasing and Saturation Rectification

BC-IHV: Conditioning the Color Space for Stable Rectified-Flow Low-Light Enhancement

Retrieval-Augmented Generation-Based Color Restoration for Low-Light Image Enhancement