AutoLumNet: A Bi-Branch Exposure-Aware Network for Low- and High-Exposure Image Enhancement

ICLR 2026 Conference Submission19722 Authors

Keywords: Image Enhancement, Exposure Correction, Deep Curve Estimation, Bi-Branch Network

TL;DR: We propose AutoLumNet, a bi-branch exposure-aware image enhancement network with a shared encoder, dual decoders, and a learnable gating module. Trained on paired low/high-light data, it adaptively corrects exposure across diverse lighting conditions

Abstract: Enhancing images captured under challenging illumination is difficult because real-world scenes often contain both severely under-exposed shadows and over-exposed highlights. Existing low-light enhancement methods primarily address under-exposure, while multi-exposure fusion requires multiple bracketed shots, which are rarely available in practice. We propose AutoLumNet, a bi-branch exposure-aware network that performs single-shot exposure correction. AutoLumNet decomposes input features into dual branches specialized for shadows and highlights, then adaptively fuses them via spatial attention. To ensure that the corrected luminance distribution aligns with natural photographs, we introduce an optimal-transport-based exposure distribution alignment mechanism, theoretically guaranteeing monotonicity and preventing spurious extrema. Training is guided by a unified exposure-aware objective combining reconstruction fidelity, distribution alignment, perceptual consistency, and regularization terms. Extensive experiments on SICE, LOL, and MIT-Adobe FiveK demonstrate that AutoLumNet achieves state-of-the-art results across under-, over-, and mixed-exposure conditions, outperforming both single-image enhancement and multi-exposure fusion baselines in PSNR/SSIM, perceptual metrics, and user studies. Our approach bridges the gap between low-light enhancement and exposure fusion, offering a principled and practical solution for real-world photography.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 19722