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

Ying Liu; Junchao Zhang; Caiyun Wu

cs.CVApr 7, 2026

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

Authors: Ying Liu, Junchao Zhang, Caiyun Wu

Organizations: 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

Abstract

Diffusion models excel at image restoration via probabilistic modeling of forward noise addition and reverse denoising, and their ability to handle complex noise while preserving fine details makes them well-suited for Low-Light Image Enhancement (LLIE). Mainstream diffusion based LLIE methods either adopt a two-stage pipeline or an auxiliary correction network to refine U-Net outputs, which severs the intrinsic link between enhancement and denoising and leads to suboptimal performance owing to inconsistent optimization objectives. To address these issues, we propose the Signal Attenuation Diffusion Model (SADM), a novel diffusion process that integrates the signal attenuation mechanism into the diffusion pipeline, enabling simultaneous brightness adjustment and noise suppression in a single stage. Specifically, the signal attenuation coefficient simulates the inherent signal attenuation of low-light degradation in the forward noise addition process, encoding the physical priors of low-light degradation to explicitly guide reverse denoising toward the concurrent optimization of brightness recovery and noise suppression, thereby eliminating the need for extra correction modules or staged training relied on by existing methods. We validate that our design maintains consistency with Denoising Diffusion Implicit Models(DDIM) via multi-scale pyramid sampling, balancing interpretability, restoration quality, and computational efficiency.

Figures & tables

Figure 1: Comparison of Architectural Workflows and Mechanisms for Three Diffusion Model-Based Low-Light Image Enhancement and Denoising Methods. Left: Workflow comparisons of three approaches: (a) two-stage LIME+LDPM, (b) DRP with a correction network, and (c) our single-stage SADM; (d) PSNR/SSIM performance on the LOLv1 dataset, where our SADM achieves the best results. Right: Comparison of forward noise addition timesteps between SADM and standard DDM, verifying the better adaptability of SADM’s signal attenuation mechanism to low-light scenarios.

Figure 2: Framework of SADM. Under the multi-scale pyramid framework, the normally exposed image $x_{0}$ is downsampled to generate $x_{0}^{z}$ , which is then contaminated with Gaussian noise to produce the noisy image $x_{t}^{z}$ . The low-light image $x_{\text{low}}$ , the image prior condition $x_{\text{dehaze}}$ (the detailed design is presented in Appendix B.5) and the position encoding $x_{\text{positionEncode}}$ (following the setting in PyDiff ( Zhou et al., 2023 ) ) are taken as joint conditional information, which is fed into the SR3Unet alongside the noisy image $x_{t}^{z}$ . Distinct loss functions are leveraged to constrain both the noise predicted by the SR3Unet and the derived target image. Ultimately, DDIM is employed to conduct sampling and inference based on random Gaussian noise.

Figure 3: Visual effect comparison of different algorithms on the LOLv1 test dataset:(a)low\high; (b)Zero-Dce; (c)UR-Net; (d)LLF; (e)PyDiff; (f)RF; (g)GSAD; (h)WBDM; (i)FD; (j)LightenD; (k)AnlightenD; (l)CLODE; (m)DRP; (n)Ours. Local magnified views are presented in the red and green boxes.

Datasets	Metrics	Zero-DCE (CVPR20)	UR-Net (CVPR22)	LLFormer (AAAI23)	PyDiff (ICCAI23)	RF (ICCV23)	GSAD (NeurIPS23)	WBDM (TOG23)	FourierD (CVPR24)	LightenD (ECCV24)	AnlightenD (TIP24)	CLODE (ICLR25)	DRP (TPAMI25)	OURS
LOLv1	PSNR $\uparrow$	14.8607	19.8416	23.6491	27.0043	25.1526	27.6405	26.3204	17.5569	19.9187	20.0503	19.6031	24.6320	28.0985
LOLv1	SSIM $\uparrow$	0.5624	0.8243	0.8163	0.8747	0.8434	0.8752	0.8445	0.6116	0.8091	0.7498	0.7175	0.8659	0.8770
LOLv2_real	PSNR $\uparrow$	18.0588	21.0933	27.7489	-	22.7938	28.6497	28.8706	16.8634	22.6514	20.4606	17.8711	23.2284	28.9338
LOLv2_real	SSIM $\uparrow$	0.5795	0.8576	0.8602	-	0.8387	0.8930	0.8759	0.6075	0.8512	0.7576	0.6830	0.8714	0.8797
LOLv2_syn	PSNR $\uparrow$	17.7564	18.2487	17.1632	-	25.6697	28.3251	29.4624	14.0556	21.6341	17.0581	17.2125	26.1412	26.7042
LOLv2_syn	SSIM $\uparrow$	0.8140	0.8209	0.7842	-	0.9282	0.9426	0.9106	0.6429	0.8664	0.7066	0.7852	0.9444	0.9224

Table 1: Quantitative evaluation results in LOLv1, LOLv2_real, LOLv2_syn

Figure 4: Visual effect comparison of different algorithms on the LOLv2_real dataset:(a)low; (b)high; (c)Zero-Dce; (d)UR-Net; (e)LLF; (f)RF; (g)GSAD; (h)WBDM; (i)FD; (j)LightenD; (k)AnlightenD; (l)CLODE; (m)DRP; (n)Ours. Local magnified views are presented in the red and green boxes.

Figure 5: Performance of our SADM on unpaired (reference-free) low-light datasets (weights pre-trained on LOLv1). This figure presents visual enhancement results for input, WBDM, DRP and SADM to validate the model’s generalization ability, with a key strength in adaptive overexposure suppression.

Groups	$\bm{k}$	${x_{0}}$ loss	Prior Condition	PSNR	SSIM
Ab.1	✗	GTmean&Percep	dehaze	20.9171	0.4547
Ab.2	✓	GTmean	dehaze	27.8724	0.8801
Ab.3	✓	Percep	dehaze	27.7354	0.8719
Ab.4	✓	GTmean&Percep	hiseq	26.1812	0.8602
Ab.5	✓	GTmean&Percep	✗	26.6877	0.8620
OURS	✓	GTmean&Percep	dehaze	28.0987	0.8770

Table 2: Ablation Study Results with Component Configuration

Figure 6: Visual Effect Comparison of Ablation Study on the LOLv1test Dataset. (a)Ablation 1 original DDPM (without $k$ ); (b)Ablation 2 only with GTmeanLoss; (c)Ablation 3 only with PerceptionLoss; (d) Ablation 4 use histogram equalization as prior condition; (e)Ablation 5 no prior condition. Local magnified views highlight the performance differences of each ablation variant.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Variation of DDPM Noise Scheduling Coefficients with Iteration Steps(value range of images is set to [-1, 1]).

Figure 8: Variation of SADM Noise Scheduling Coefficients with Iteration Steps(value range of images is set to [0, 1]).

Figure 9: Variation of SADM Noise Scheduling Coefficients with 0.99.

Figure 10: Variation of SADM Noise Scheduling Coefficients with 0.9999.

Figure 11: Visual Comparison of Dehazing and Histogram Equalization Preprocessing With RGB Channel Distribution Analysis for Diffusion Based Low Light Image Enhancement.

Figure 12: Visual effect comparison of different algorithms on the LOLv1test dataset:(a)low; (b)high; (c)Zero-Dce; (d)UR-Net; (e)LLF; (f)PyDiff; (g)RF; (h)GSAD; (i)WBDM; (j)FD; (k)LightenD; (l)AnlightenD; (m)CLODE; (m)DRP; (n)Ours. Local magnified views are presented in the red and green boxes.

Figure 13: Visual effect comparison of different algorithms on the LOLv2_real dataset:(a)low; (b)high; (c)Zero-Dce; (d)UR-Net; (e)LLF; (f)RF; (g)GSAD; (h)WBDM; (i)FD; (j)LightenD; (k)AnlightenD; (l)CLODE; (m)DRP; (n)Ours. Local magnified views are presented in the red and green boxes.

Figure 14: Visual effect comparison of different algorithms on the LOLv2_syn dataset:(a)low; (b)high; (c)Zero-Dce; (d)UR-Net; (e)LLF; (f)RF; (g)GSAD; (h)WBDM; (i)FD; (j)LightenD; (k)AnlightenD; (l)CLODE; (m)DRP; (n)Ours.

Figure 15: Visual effect comparison of different algorithms on the LOLv2_syn dataset.

Figure 16: SADM’s visual results on the unpaired dataset.

Explore similar work

P-PatchDiff: Progressive Patch Diffusion Models for Low-light Image Enhancement
Sep 1, 2026Ruoyu Guo, Haonan Zhong, Maurice Pagnucco +1Diffusion Models for Image Restoration Low-Light Image Enhancement
School of Computer Science and Engineering, University of New South Wales, Sydney, Australia.
IDM-Net: A Lightweight Illumination-Decoupled Modulation Network for Low-Light Image Enhancement
Sep 25, 2026Cheng-Yen Hsiao, Jing-Ming GuoLow-Light Image Enhancement
Department of Electrical Engineering National Taiwan University of Science and Technology Taipei, Taiwan
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

Sep 1, 2026J/K move · Enter open · S save

P-PatchDiff: Progressive Patch Diffusion Models for Low-light Image Enhancement

Ruoyu Guo, Haonan Zhong, Maurice Pagnucco, Yang Song

School of Computer Science and Engineering, University of New South Wales, Sydney, Australia.

Diffusion Models for Image Restoration Low-Light Image Enhancement Diffusion Model Inference Acceleration Diffusion Models

Open paper AskPDF arXiv

Recent advancements in low-light image enhancement have leveraged diffusion models for their strong ability to generate perceptually realistic, detailed images. Patch diffusion models further offer a promising solution to size-agnostic image restoration while improving efficiency. However, existing methods typically rely on small, fixed patches (e.g., 64

\times

64) that cannot capture image-level brightness context, whereas enlarging the receptive field improves brightness and colour estimation but substantially increases computational cost. Moreover, low-light images often exhibit uneven brightness across regions, making it necessary to ensure that locally enhanced patches remain visually coherent when combined into the full image. To address these limitations, we propose P-PatchDiff, a scalable progressive patch diffusion framework for low-light image enhancement that dynamically adjusts patch size throughout the denoising process, enabling a gradual shift from local to global views. A Multi-Patch Alignment strategy is also introduced to normalise features across varying patch scales using an estimated global brightness proxy. Rather than pursuing pixel-level reconstruction accuracy, P-PatchDiff focuses on scalability and coherent brightness across the whole image, allowing the model to perceive multi-scale information and better enhance regions with varying brightness. We empirically demonstrate that P-PatchDiff effectively enhances images ranging from 400

\times

600 to 4K and is 80

\times

faster than existing patch diffusion models while using less than 9GB of memory. The code is available at https://github.com/RuoyuGuo/P-PatchDiff.

Abstract

Explore similar work

P-PatchDiff: Progressive Patch Diffusion Models for Low-light Image Enhancement

IDM-Net: A Lightweight Illumination-Decoupled Modulation Network for Low-Light Image Enhancement

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