More Works

Anchor then Polish for Low-light Enhancement

Tianle Du,† Mingjia Li,† Hainuo Wang, Xiaojie Guo*
School of Software, Tianjin University, Tianjin, China
†Equal Contribution
*Corresponding Author
{dutianle, mingjiali, hainuo}@tju.edu.cn, xj.max.guo@gmail.com
Paper Code arXiv
Teaser image for Anchor then Polish for Low light Enhancement

We uncover a pronounced energy imbalance in LLIE, showing that over 90% of the residual MSE arises from linear global illumination shifts, which can be effectively corrected using a simple global anchoring operator with only 12 degrees of freedom.

Abstract

Low-light image enhancement is challenging due to entangled degradations, mainly including poor illumination, color shifts, and texture interference. Existing methods often rely on complex architectures to address these issues jointly but may overfit simple physical constraints, leading to global distortions. This work proposes a novel anchor-then-polish (ATP) framework to fundamentally decouple global energy alignment from local detail refinement. First, macro anchoring is customized to (greatly) stabilize luminance distribution and correct color by learning a scene-adaptive projection matrix with merely 12 degrees of freedom, revealing that a simple linear operator can effectively align global energy. The macro anchoring then reduces the task to micro polishing, which further refines details in the wavelet domain and chrominance space under matrix guidance. A constrained luminance update strategy is designed to ensure global consistency while directing the network to concentrate on fine-grained polishing. Extensive experiments on multiple benchmarks show that our method achieves state-of-the-art performance, producing visually natural and quantitatively superior low-light enhancements.

Method Overview

First research result visualization

(a) Architecture of the proposed ATP framework. (b) Global Energy Anchoring Module (GEAM). (c) Detail Polishing Module (DPM).

Global Matrix Analysis

Subfigure A visual result

(a) Matrix structure

Matrix structure of GEAM prediction on LOLv2-Real. Here, \(\rho\) indicates a monotonic relationship between matrix elements and input luminance. The closer \(\left|\rho\right|\) is to 1, the stronger the relationship.

(b) Matrix Guidance

Matrix Guidance performs adaptive feature modulation based on degradation. In moderate low-light, subtle gating preserves textures and limits noise. In severe cases, stronger modulation prioritizes structural features and suppresses noise-dominated responses, ensuring effective denoising and detail recovery.

Subfigure B visual result
Subfigure C visual result

(c) Matrix Effect and CLU

GEAM achieves SOTA via a $3 \times 4$ affine matrix, ensuring superior luminance consistency. CLU stabilizes this alignment by suppressing training oscillations, maintaining a consistent optimization objective that allows the refinement stage to focus exclusively on high-fidelity texture recovery.

Quantitative comparison

Visual Comparison

Paper poster

Visual comparison on representative low-light scenes across multiple benchmarks. ATP restores more natural brightness, color fidelity, and fine details while effectively suppressing noise and artifacts. Please refer to our paper for more visualization results.

BibTeX

@misc{du2026atp,
      title={Anchor then Polish for Low-light Enhancement}, 
      author={Tianle Du and Mingjia Li and Hainuo Wang and Xiaojie Guo},
      year={2026},
      eprint={2603.15472},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.15472}, 
}