Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement

Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, Xinghao Ding

TL;DR

Real-world low-light image enhancement (LIE) suffers from scarce degraded/ground-truth pairs and unknown, complex degradations. The paper presents AGLLDiff, a training-free, unsupervised diffusion-based framework that guides sampling using high-quality image attributes—exposure, structure, and color—without modeling the degradation, enabling robust enhancement across real-world data. It deploys a dynamic attribute-guided sampling strategy with losses on exposure, phase, and Retinex-based color constraints, and demonstrates superior performance to unsupervised baselines and competitive results against some supervised methods on eight challenging LIE benchmarks. The work offers practical impact by delivering high-fidelity LIE without the need for paired data or degradation modeling, while highlighting avenues for faster sampling and expansion to additional restoration tasks.

Abstract

Existing low-light image enhancement (LIE) methods have achieved noteworthy success in solving synthetic distortions, yet they often fall short in practical applications. The limitations arise from two inherent challenges in real-world LIE: 1) the collection of distorted/clean image pairs is often impractical and sometimes even unavailable, and 2) accurately modeling complex degradations presents a non-trivial problem. To overcome them, we propose the Attribute Guidance Diffusion framework (AGLLDiff), a training-free method for effective real-world LIE. Instead of specifically defining the degradation process, AGLLDiff shifts the paradigm and models the desired attributes, such as image exposure, structure and color of normal-light images. These attributes are readily available and impose no assumptions about the degradation process, which guides the diffusion sampling process to a reliable high-quality solution space. Extensive experiments demonstrate that our approach outperforms the current leading unsupervised LIE methods across benchmarks in terms of distortion-based and perceptual-based metrics, and it performs well even in sophisticated wild degradation.

AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement

TL;DR

Real-world low-light image enhancement (LIE) suffers from scarce degraded/ground-truth pairs and unknown, complex degradations. The paper presents AGLLDiff, a training-free, unsupervised diffusion-based framework that guides sampling using high-quality image attributes—exposure, structure, and color—without modeling the degradation, enabling robust enhancement across real-world data. It deploys a dynamic attribute-guided sampling strategy with losses on exposure, phase, and Retinex-based color constraints, and demonstrates superior performance to unsupervised baselines and competitive results against some supervised methods on eight challenging LIE benchmarks. The work offers practical impact by delivering high-fidelity LIE without the need for paired data or degradation modeling, while highlighting avenues for faster sampling and expansion to additional restoration tasks.

Abstract

Existing low-light image enhancement (LIE) methods have achieved noteworthy success in solving synthetic distortions, yet they often fall short in practical applications. The limitations arise from two inherent challenges in real-world LIE: 1) the collection of distorted/clean image pairs is often impractical and sometimes even unavailable, and 2) accurately modeling complex degradations presents a non-trivial problem. To overcome them, we propose the Attribute Guidance Diffusion framework (AGLLDiff), a training-free method for effective real-world LIE. Instead of specifically defining the degradation process, AGLLDiff shifts the paradigm and models the desired attributes, such as image exposure, structure and color of normal-light images. These attributes are readily available and impose no assumptions about the degradation process, which guides the diffusion sampling process to a reliable high-quality solution space. Extensive experiments demonstrate that our approach outperforms the current leading unsupervised LIE methods across benchmarks in terms of distortion-based and perceptual-based metrics, and it performs well even in sophisticated wild degradation.
Paper Structure (11 sections, 12 equations, 9 figures, 2 tables, 1 algorithm)

This paper contains 11 sections, 12 equations, 9 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Preliminaries of Diffusion Models
  5. Overview of the AGLLDiff Framework
  6. Attribute Guidance
  7. Experiments
  8. Implementation and Datasets
  9. Comparison with the State-of-the-Art
  10. Ablation study
  11. Conclusion and Limitations

Figures (9)

  • Figure 1: Motivation of our AGLLDiff. (a) represents the data distribution of normal-light samples and degraded samples. It is evident that degradation significantly deviates from normal-light samples. (b) conceptually illustrate the geometries of the proposed attribute guidance sampling algorithm. It shows that, given the initial latent, which lies in the low-probability region, attribute guidance guides the latent to move towards its vicinal high-probability region. (c) presents that imposing gaussian noise on the degraded sample and its corresponding reference sample makes the distributions between them less distinguishable.
  • Figure 2: The overall framework of our proposed AGLLDiff.
  • Figure 3: Statistics and Visualization. (a) The average exposure values of the low- and normal-light subsets of the five LIE datasets. (b) Visualization of the structure of low- and normal-light images by the Canny operator. (c) Histogram of the color distribution of the low- and normal-light images.
  • Figure 4: (a) Visualization of the spatially variant exposure maps. Based on Eq. \ref{['eq8']}, we automatically assign the underexposed regions large exposure values (light gray) and wellexposed/overexposed regions small exposure values (dark gray). (b) Visualization of the phase-only reconstruction image (PCI) in the spatial domain. We applied an inverse discrete Fourier transform to the phase of the low/normal-light image to obtain the phase-only reconstruction image. That means the amplitude of low/normal-light image is set to 1. (c) Visualization of the Retinex decomposition. We employ a pre-trained decomposition network, RNet, to decompose the input into a reflectance map $R$ and an illumination map $L$.
  • Figure 5: Visual comparisons of various LIE methods on SICE. The proposed method achieves visually pleasing results in terms of brightness, color, contrast, and naturalness.
  • ...and 4 more figures