More Works

Learning Structural Illumination for Unsupervised Low-Light Enhancement

Tianle Du,1 Peiyuan He,1 Hainuo Wang,1 Tianxiu Yu,2 Xiaojie Guo1,*
1Tianjin University    2Dunhuang Academy Preprint
*Corresponding Author
{dutianle, peiyuan_he, hainuo}@tju.edu.cn, xj.max.guo@gmail.com
Paper Code Supplement
Loading figure...

Figure 1. RISE consistently delivers a superior PSNR-FLOPs trade-off across diverse benchmark brightness distributions while explicitly separating relative illumination structure from absolute exposure.

Abstract

Existing unsupervised low-light image enhancement (LLIE) methods often estimate illumination directly from the entire low-light input, without separating its spatially varying illumination pattern, termed relative illumination structure, from the absolute exposure level or preventing unreliable low signal-to-noise ratio regions from biasing the estimate. Moreover, fixed exposure targets impose a scene-agnostic enhancement criterion, limiting adaptation across diverse lighting conditions. Inspired by the spatial propagation of light, we propose a Relative Illumination Structure Estimation (RISE) framework that decouples relative illumination structure from absolute exposure and infers it from reliable bright regions, enabling interpretable and robust enhancement. For scene-adaptive exposure adjustment, we further propose a Dual-Metering Exposure Reference derived from each input, allowing RISE to adapt the enhancement strength to individual scenes and generalize across diverse lighting conditions. Extensive benchmark and real-world generalization experiments show that RISE achieves state-of-the-art performance among unsupervised LLIE methods while producing visually natural results.

Structural Illumination Analysis

Unsupervised LLIE faces two closely related difficulties. First, severely underexposed regions have a low signal-to-noise ratio, yet most methods treat the entire image as equally reliable when estimating illumination. This can make the model brighten noisy shadows incorrectly while simultaneously overexposing highlights. Second, existing representations usually entangle the scene's spatial illumination pattern with its global exposure level, making their individual roles difficult to identify or control.

RISE starts from a different view: illumination contains a relative structure that describes how brightness varies across the scene and an absolute exposure that controls the overall output level. Reliable bright regions lie on stronger light-propagation paths, so their relationships to darker regions provide observable cues for recovering this structure.

Loading figure...

Learning from reliable illumination cues

Light strength changes spatially because of distance, occlusion, absorption, and other scene factors. Explicitly reconstructing this physical propagation from a single image is ill-posed, but the relative brightness relationships remain observable.

RISE therefore selects spatially dispersed bright patches as Key Illumination Cues (KICs). Instead of allowing unreliable dark observations to dominate the estimate, these KICs act as anchors from which illumination relationships are propagated to the rest of the scene.

Loading figure...

Stable structure, adaptive exposure

Different exposures of the same scene preserve nearly identical relative parameters, while the independently predicted exposure offset changes with input brightness. This allows RISE to enhance underexposed inputs and also suppress excessive exposure without altering the underlying spatial structure.

The substantially lower consistency between different scenes confirms that the relative parameters describe scene-specific illumination organization rather than a fixed enhancement style.

Method Overview

Given a low-light input, RISE first removes global brightness from its pooled luminance to obtain a relative illumination map. It then selects a small set of bright, spatially dispersed KICs and predicts their target luminance. A relation branch connects every image patch to these reliable anchors in both luminance and spatial domains, forming a structured illumination field.

In parallel, a separate exposure branch predicts one image-level offset that controls global brightness. The final correction is composed in log-gain space, so relative structure and absolute exposure remain explicitly decoupled and can respond independently to scene content and input exposure.

Loading figure...

RISE jointly encodes image appearance and relative illumination, estimates a KIC-anchored structural field, and independently controls the absolute exposure of the enhanced image.

01

Relative illumination map

Patch luminance is normalized by its image-wide mean and mapped to log space. This removes global brightness while retaining signed local illumination differences.

02

KIC relation modeling

Each patch is related to the KICs through luminance discrepancy and a bounded spatial attenuation prior, while scene statistics provide global structural context.

03

Decoupled exposure control

Separate predictors estimate structural parameters and the exposure offset, leaving the structure stable while global brightness adapts to each input.

Scene-adaptive exposure supervision

Fixed exposure targets encourage different scenes to converge toward the same brightness preference. The proposed Dual-Metering Exposure Reference (DMER) instead derives its training reference from each input. An absolute meter provides conservative global calibration, while a relative meter diffuses reliable luminance cues to describe spatial illumination demand. Blending both references provides adaptive supervision without paired normal-light images or training-set brightness statistics.

Loading figure...

The absolute meter estimates global exposure demand, while the relative meter follows spatial illumination variation to produce an input-dependent final reference.

Quantitative Comparison

Best    Second best    RISE uses the same LSRW-Nikon-trained checkpoint across datasets; RISE† is trained per benchmark.

Visual Comparison

Loading figure...

RISE produces natural exposure correction and preserves color fidelity across LOL and UHD-LL.

Loading figure...

Without additional tuning, RISE generalizes to real-world iPhone 16 Pro captures while preserving a broader dynamic range.

BibTeX

@misc{du2026rise,
  title     = {Learning Structural Illumination for Unsupervised Low-Light Enhancement},
  author    = {Du, Tianle and He, Peiyuan and Wang, Hainuo and Yu, Tianxiu and Guo, Xiaojie},
  note      = {Preprint},
  year      = {2026}
}