Diffusion-Based Low-Light Image Enhancement with Color and Luminance Priors

Xuanshuo Fu; Lei Kang; Javier Vazquez-Corral

cs.CVFeb 27, 2026

Diffusion-Based Low-Light Image Enhancement with Color and Luminance Priors

Authors: Xuanshuo Fu, Lei Kang, Javier Vazquez-Corral

Organizations: Computer Vision Center, Cerdanyola del Vallès, Spain and Universitat Autònoma de Barcelona, Cerdanyola del Vallès, Spain

Abstract

Low-light images often suffer from low contrast, noise, and color distortion, degrading visual quality and impairing downstream vision tasks. We propose a novel conditional diffusion framework for low-light image enhancement that incorporates a Structured Control Embedding Module (SCEM). SCEM decomposes a low-light image into four informative components including illumination, illumination-invariant features, shadow priors, and color-invariant cues. These components serve as control signals that condition a U-Net-based diffusion model trained with a simplified noise-prediction loss. Thus, the proposed SCEM equipped Diffusion method enforces structured enhancement guided by physical priors. In experiments, our model is trained only on the LOLv1 dataset and evaluated without fine-tuning on LOLv2-real, LSRW, DICM, MEF, and LIME. The method achieves state-of-the-art performance in quantitative and perceptual metrics, demonstrating strong generalization across benchmarks. https://casted.github.io/scem/.

Figures & tables

Panel of Figure 1 — Fig. 1 : Quantitative comparisons with state-of-the-art methods. (a) presents numerical scores for PSNR, SSIM, LPIPS, and FID on 3 datasets: LOLv1, LOLv2-real, and LSRW, which contain groundtruth normal-light images. (b) presents numerical scores for NIQE, BRISQUE and PI on 3 datasets: DICM, MEF, and LIME, which contain only low-light images without groundtruth normal-light counterparts. Note that our proposed methods are trained only on the LOLv1 training set and are directly evaluated on the remaining datasets. To enable intuitive comparison in the radar plot, we normalized all metrics and inverted those where lower is better (e.g., LPIPS, FID), so that higher values consistently indicate better performance.

Fig. 2 : The proposed architecture of SCEM equiped diffusion model for low-light image enhancement. During training, we extract four types of features from the input low-light image $\mathcal{I}$ : illumination $T_{ref}$ , color invariance features $\Phi(x)$ , illumination-invariant features $R_{c}$ , and shadow priors $S_{3ch}$ . These features, along with the original low-light image, are concatenated with the randomly chosen $t$ -th noised image $X_{t}$ to form the input $X_{t}^{\prime}$ for the denoising training process. The diffusion model then generates the enhanced image $\hat{X}_{0}$ . During inference, Gaussian noise $X_{T}$ is concatenated with the same set of extracted features and the original low-light image $I$ .

Methods	Reference	LOL v1				LOLv2-real				LSRW
Methods	Reference	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
NPE [ 27 ]	TIP’ 13	16.970	0.484	0.400	104.057	17.333	0.464	0.396	100.025	16.188	0.384	0.440	90.132
SRIE [ 28 ]	CVPR’ 16	11.855	0.495	0.353	88.728	14.451	0.524	0.332	78.834	13.357	0.415	0.399	69.082
LIME [ 29 ]	TIP’ 16	17.546	0.531	0.387	117.892	17.483	0.505	0.428	118.171	17.342	0.520	0.471	75.595
RetinexNet [ 5 ]	BMVC’ 18	16.774	0.462	0.417	126.266	17.715	0.652	0.436	133.905	15.609	0.414	0.454	108.350
DSLR [ 30 ]	TMM’ 20	14.816	0.572	0.375	104.428	17.000	0.596	0.408	114.306	15.259	0.441	0.464	84.930

TABLE I : Quantitative evaluation on three datasets: LOLv1 [ 5 ] , LOLv2-real [ 25 ] , and LSRW [ 26 ] . All values are formatted with three decimal places. Higher PSNR and SSIM indicate better performance, while lower LPIPS and FID are preferable. Note that methods are trained only on the LOLv1 training set and are directly evaluated on the remaining datasets

Fig. 3 : Visual comparisons of our approach with competing methods. The input image is from datasets LOLv1, LOLv2-real, LSRW for the first, second and third rows, respectively.

Methods	DICM			MEF			LIME
Methods	NIQE $\downarrow$	BRI $\downarrow$	PI $\downarrow$	NIQE $\downarrow$	BRI $\downarrow$	PI $\downarrow$	NIQE $\downarrow$	BRI $\downarrow$	PI $\downarrow$
$\text{LIME}_{\text{\cite[cite]{[\@@bibref{}{guo2016lime}{}{}]}}}$	4.476	27.375	4.216	4.744	39.095	5.160	5.045	32.842	4.859
$\text{DRBN}_{\text{\cite[cite]{[\@@bibref{}{yang2020fidelity}{}{}]}}}$	4.369	30.708	3.800	4.869	44.669	4.711	4.562	34.564	3.973
$\text{Zero-DCE}_{\text{\cite[cite]{[\@@bibref{}{guo2020zero}{}{}]}}}$	3.951	23.350	3.149	3.500	29.359	2.989	4.379	26.054	3.239
$\text{MIRNet}_{\text{\cite[cite]{[\@@bibref{}{zamir2020learning}{}{}]}}}$	4.021	22.104	3.691	4.202	34.499	3.756	4.378	28.623	3.398
$\text{EnlightenGAN}_{\text{\cite[cite]{[\@@bibref{}{jiang2021enlightengan}{}{}]}}}$	3.832	19.129	3.256	3.556	26.799	3.270	4.249	22.664	3.381

TABLE II : Quantitative comparison of different methods on the DICM [ 44 ] , MEF [ 45 ] , and LIME [ 29 ] datasets. The best results are highlighted in bold and the second-best in underlined , and BRI represents BRISQUE [ 51 ] .

SCEM	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$-$	22.220	0.810	0.220
✓	26.947	0.921	0.071

TABLE III : Ablation study for our proposed SCEM module on LOLv1.

Fig. 4 : Ablation study comparing PSNR for different input configurations: (1) low-light image $\mathcal{I}$ only, (2) $\mathcal{I}$ with illumination $T_{ref}$ , (3) $\mathcal{I}$ with color-invariant features $\Phi(x)$ , (4) $\mathcal{I}$ with illumination-invariant features $R_{c}$ , and (5) $\mathcal{I}$ with shadow priors $S_{3ch}$ .

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
Single-Stage Signal Attenuation Diffusion Model for Low-Light Image Enhancement and Denoising
Apr 7, 2026Ying Liu, Junchao Zhang, Caiyun WuLow-Light Image Enhancement Diffusion Models for Image Restoration
School of Aeronautics and Astronautics, Central South University, Changsha 410083, China · Hunan Provincial Key Laboratory of Optic-Electronic Intelligent Measurement and Control, Changsha 410083, China
DARD: Zero-Shot Degradation-Aware Retinex-Guided Diffusion for Low-Light Image Enhancement
Aug 29, 2026Wenjie Cai, Yuezhe Yang, Jianyang Xia +2Image Enhancement Low-Light Image Enhancement
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

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

Single-Stage Signal Attenuation Diffusion Model for Low-Light Image Enhancement and Denoising

DARD: Zero-Shot Degradation-Aware Retinex-Guided Diffusion for Low-Light Image Enhancement