Towards Scale-Aware Low-Light Enhancement via Structure-Guided Transformer Design

Abstract

We present a novel Transformer-based framework for low-light image enhancement, Towards Scale-Aware Low-Light Enhancement via Structure-Guided Transformer Design. Our model is built upon a U-Net-style encoder–decoder architecture, where we introduce a customized Hybrid Structure-Guided Feature Extractor (HSGFE) at each scale. The HSGFE integrates three key components: (1) a Dilated Residual Dense Block (DRDB) for effective feature refinement, (2) a Structure-Guided Transformer Block (SGTB) that incorporates structural priors to preserve edges and suppress noise, and (3) a Semantic-Aligned Scale-Aware Module (SAM) to handle multi-scale variations. This design enables our network to enhance low-light images while maintaining structural integrity and reducing color distortion. Extensive experiments show that our method achieves state-of-the-art performance in both quantitative metrics and visual quality. Our approach also achieve top-tier results on standard LLIE benchmarks and ranked second in the NTIRE 2025 Low-Light Image Enhancement Challenge.