Source-linked AI summary
Diff-Retinex: Rethinking Low-light Image Enhancement with A Generative Diffusion Model
Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, Jiayi Ma
TL;DR
Low-light enhancement must handle degradations including noise, low contrast, color deviation, and lost scene structures that existing fitting-based methods struggle to recover. Diff-Retinex combines Retinex decomposition with conditional diffusion generation, and experiments report strong performance with subtle-detail completion and inference restoration. Its pixel-wise error metrics do not consistently dominate, reflecting a trade-off between generation effects and stricter pixel constraints.
Problem
Low-light images contain diverse degradations and missing scene content, while existing methods struggle to repair lost information.
Method
Diff-Retinex combines Transformer-based Retinex decomposition into illumination and reflectance maps with multi-path conditional diffusion networks for their degradations.
Results
Experiments report excellent performance, including subtle-detail completion and inference restoration, with qualitative results showing generated textures and completed missing scenes.
Takeaways & Limitations
Generative diffusion enables low-light enhancement to address missing scene contents and subtle details beyond restoration of the original low-quality information.
Takeaways & Limitations
Diff-Retinex does not establish dominance on pixel-wise error metrics such as PSNR, because stronger pixel constraints can weaken generation effects.
Abstract
from arXiv · showhide
In this paper, we rethink the low-light image enhancement task and propose a physically explainable and generative diffusion model for low-light image enhancement, termed as Diff-Retinex. We aim to integrate the advantages of the physical model and the generative network. Furthermore, we hope to supplement and even deduce the information missing in the low-light image through the generative network. Therefore, Diff-Retinex formulates the low-light image enhancement problem into Retinex decomposition and conditional image generation. In the Retinex decomposition, we integrate the superiority of attention in Transformer and meticulously design a Retinex Transformer decomposition network (TDN) to decompose the image into illumination and reflectance maps. Then, we design multi-path generative diffusion networks to reconstruct the normal-light Retinex probability distribution and solve the various degradations in these components respectively, including dark illumination, noise, color deviation, loss of scene contents, etc. Owing to generative diffusion model, Diff-Retinex puts the restoration of low-light subtle detail into practice. Extensive experiments conducted on real-world low-light datasets qualitatively and quantitatively demonstrate the effectiveness, superiority, and generalization of the proposed method.
1. Introduction
Low-light images suffer from multiple degradations, especially lost scene structures and information. Diff-Retinex addresses these issues by combining Retinex decomposition with conditional diffusion-based generation.
- Motivation: Low-light images exhibit noise, low contrast, color deviation, and lost scene structures that reduce visual quality and information.The paper identifies scene-structure loss as the most difficult degradation.
- Motivation: Existing methods often fit or denoise distorted scenes but cannot repair missing content, and may worsen information distortion.The paper contrasts this limitation with its generative formulation of enhancement.
- Diff-Retinex: Diff-Retinex formulates low-light enhancement as Retinex decomposition plus conditional image generation.This combines a physically explainable model with a generative network.
- Diff-Retinex: The Retinex Transformer decomposition network decomposes images into illumination and reflectance maps while exploiting attention and layer dependence.The network is designed to improve decomposition applicability, including for high-resolution images.
- Diff-Retinex: Multi-path diffusion networks separately address dark illumination, noise, color deviation, and missing scene contents in the decomposed components.The approach aims to compensate for content loss and color deviation rather than only enhance existing low-quality information.
2. Related Work
Retinex-based methods decompose low-light images into illumination and reflectance components, while generative approaches model enhancement as image generation. Existing methods include traditional filtering and priors, deep CNN-based decomposition, and direct low-light-to-normal-light generation.
- Retinex-based methods: Retinex methods decompose images into illumination and reflectance maps for low-light enhancement.
- Traditional approaches: Traditional approaches estimate or adjust illumination and reflectance using filters, channel maxima, structural priors, sequence decomposition, or gamma transforms.
- Deep learning-based approaches: Deep learning methods combine Retinex decomposition with learned adjustment, using CNNs to map decomposition and enhancement components.
- Deep learning-based approaches: CNN-based decomposition cannot fully exploit global information because of convolutional limitations and faces difficult loss-function design.
- Generative approaches: Generative approaches condition image generation on low-light inputs to produce corresponding normal-light images, including direct mapping methods such as EnlightenGAN.
3. Methodology
Diff-Retinex decomposes low-light images into illumination and reflectance, then adjusts these components with multi-path diffusion generation. Its Transformer decomposition network uses attention-based design choices intended to preserve information while reducing high-resolution computation.
- Retinex decomposition: The Retinex framework decomposes an input image into illumination and reflectance maps before enhancement.Reflectance represents scene content, while illumination represents lighting conditions and is expected to be locally smooth.
- Retinex decomposition: The decomposition objective combines reconstruction, reflectance consistency, and illumination smoothness losses.These losses respectively support image reconstruction, consistency across lighting conditions, and piece-wise smooth illumination.
- Network architecture: The Transformer Decomposition Network has separate reflectance and illumination branches, with a multi-stage Transformer encoder-decoder in the reflectance branch.The reflectance branch is designed to improve recovery performance and information retention.
- Network architecture: MDLA computes attention across channels and uses depth-wise convolutions to reduce the quadratic overhead of standard Transformer attention for high-resolution images.It aggregates information with 1 × 1, 3 × 3, 5 × 5, and 7 × 7 convolutions before reshaping features for attention.
- Diffusion generation adjustment: The diffusion generation adjustment uses separate reflectance and illumination paths to reconstruct normal-light Retinex component distributions.The process follows DDPM-style forward diffusion, which progressively adds Gaussian noise, and reverse diffusion guided by conditional images.
4. Experiment
Experiments on real-world low-light datasets compare Diff-Retinex with traditional and learning-based methods, showing strong qualitative restoration and broad quantitative performance. Additional analyses examine decomposition and diffusion generation, while discussion acknowledges a pixel-wise error limitation.
- Qualitative Comparison: Diff-Retinex restores missing textures and scene details that competing methods fail to recover in qualitative comparisons.Examples include coarse-grained ground texture, a diving platform, and a handrail.
- Qualitative Comparison: Diff-Retinex provides better illumination and color fidelity than competing methods in the reported qualitative examples.The paper reports that its colors are closest to ground truth, while several competitors show deviations.
- Quantitative Comparison: Diff-Retinex achieves strong quantitative results on LOL and VE-LOL-L using FID, LPIPS, BIQI, LOE, and PI, including comprehensive best performance on VE-LOL-L.On LOL, it is especially advantageous on FID and LPIPS and is slightly lower than LLFormer on LOE; on VE-LOL-L, it achieves the comprehensive best metric performance.
- Quantitative Comparison: On DICM, Diff-Retinex demonstrates competitiveness, while the experiments also report PSNR and SSIM comparisons on LOL.The evaluation uses multiple datasets and includes cross-dataset generalization testing.
- Ablation and Analysis: The ablation analyses visualize TDN decomposition, RDA and IDA diffusion recovery, and reflectance restoration against other low-light enhancement methods.Reflectance maps are used because they contain substantial color and texture information and are sensitive to visual perception.
- Discussion: Diff-Retinex does not dominate pixel-wise error metrics such as PSNR, reflecting a reported trade-off between stricter pixel constraints and generation quality.The discussion presents improved pixel-level performance through diffusion models as a desirable direction.
5. Conclusion
The paper concludes that Diff-Retinex reframes low-light enhancement as decomposition and image generation. It adaptively decomposes images, addresses multiple degradations with diffusion models, and enables subtle-detail completion and inference restoration.
- Conclusion: Diff-Retinex formulates low-light enhancement as a paradigm combining Retinex decomposition and image generation.The conclusion presents this formulation as the paper's central framework.
- Conclusion: Diff-Retinex adaptively decomposes images into illumination and reflectance maps and addresses various degradations through generative diffusion models.The conclusion links the decomposition and diffusion stages within the proposed model.
- Conclusion: Experimental results show excellent performance and make subtle-detail completion and inference restoration practical for low-light enhancement.This is the conclusion's stated outcome for the proposed generative model.