RePriFusion: Retinex-Based Low-Light Image Enhancement via Prior-Specific Fusion

Low-light image enhancement increasingly benefits from physical and pretrained visual priors. However, existing methods often rely on a single prior or integrate multiple representations through uniform fusion, limiting the effective use of their complementary information. We propose $\text{RePriFusion}$, a low-light image enhancement framework that integrates learned retinex-based prior features and vision foundation model features through selected restoration stages and prior-specific fusion modules. First, we design a cascaded Retinex estimator to progressively extract illumination and reflectance prior features and constructs an intermediate reconstruction. Second, we design prior-specific fusion modules that introduce illumination, reflectance, and the foundation model priors into selected U-Net stages using dedicated fusion operators tailored to their restoration roles. Finally, mixed-dataset training and noise augmentation are further adopted to improve robustness to diverse low-light distributions. Experiments on five paired LLIE benchmarks and eight unseen datasets show strong enhancement quality and zero-shot generalization.