LLDiffusion: Learning degradation representations in diffusion models for low-light image enhancement

Current deep learning methods for low-light image enhancement typically rely on pixel-wise mappings using paired data, often overlooking the specific degradation factors inherent to low-light conditions, such as noise amplification, reduced contrast, and color distortion. This oversight can result in suboptimal performance. To address this limitation, we propose a degradation-aware learning framework that explicitly integrates degradation representations into the model design. We introduce LLDiffusion, a novel model composed of three key modules: a Degradation Generation Network (DGNET), a Dynamic Degradation-Aware Diffusion Module (DDDM), and a Latent Map Encoder (E). This approach enables joint learning of degradation representations, with the pre-trained Encoder (E) and DDDM effectively incorporating degradation and image priors into the diffusion process for improved enhancement. Extensive experiments on public benchmarks show that LLDiffusion outperforms state-of-the-art low-light image enhancement methods quantitatively and qualitatively. The source code and pre-trained models will be available at https://github.com/TaoWangzj/LLDiffusion.