A multimodal intrinsics-guided thermal-aware framework for RGB low-light image enhancement

Low-light image enhancement is crucial in situations where visible sensors might suffer from severe noise and information loss (e.g., nighttime surveillance). Recent approaches

investigate auxiliary modalities invariant to illumination to improve the performance, such as thermal infrared imaging. We propose a Multimodal Intrinsics-Guided Framework that integrates RGB and thermal data to reconstruct well-lit images. Our method utilizes a two-stage pipeline: first, we employ an intrinsic decomposition strategy to separate reflectance and shading components through knowledge distillation, where a teacher network guides a student model in reconstructing consistent intrinsic components; then, a refinement stage restores fine structures and visual details. We train the proposed model on synthetic data from HDRT dataset and demonstrate strong generalization to real-world benchmarks such as LLVIP and V-TIEE, outperforming state-of-the-art methods in most evaluation metrics. Code is available at: https://github.com/simonemelc/TIRGlow