IADFormer: Intensity-Aligned Differential Transformer for Low-Light Image Enhancement

Abstract

Low-light images usually suffer from spatially uneven illumination and exposure-dependent feature degradation, where weak structural and chromatic cues are easily masked by dominant intensity variations. Existing Transformer-based methods improve feature interaction, but often overlook the ambiguity caused by similar low-intensity appearances in dark regions and their interference in attention. To address these issues, we propose IADFormer, an intensity-aligned differential Transformer for low-light image enhancement. We design an Intensity-Aligned Differential Self-Attention (IDSA) module to model feature dependencies in an intensity-aligned representation and contrast complementary attention responses, enabling the network to weaken redundant low-intensity responses and better preserve weak structural and chromatic cues. To handle spatially non-uniform illumination, we introduce a Local Brightness Prior Modulation (LBPM) branch that uses local brightness priors from the input image to adaptively guide feature restoration. In addition, we introduce an Intensity-Aligned Correlation Loss (IACL) to align global intensity and inter-channel color relationships. Extensive experiments demonstrate that IADFormer achieves competitive performance.