FacDNet: A low-rank factorized diffusion network with dual-U compensated attention for low-light enhancement

Low-light image enhancement (LLIE) is to improve visibility and preserve structural integrity in poorly illuminated scenes. Given the presence of frequent and non-uniform illumination degradation, existing methods often struggle to balance structural consistency and detail fidelity. Motivated by diffusion models that excel at generating fine details via iterative denoising, we propose a low-rank factorized diffusion network with dual-U compensated attention (FacDNet) which establishes an information flow across the feature and spatial domains for structure-aware detail enhancement. Specifically, in the feature domain, we design a low-rank factorized diffusion architecture that decomposes original image contents into principal components during forward decoupling. In the reverse process, these components are recomposed with controllable noise injection to reconstruct them back into the image space, supporting structural modeling and detail refinement. In the spatial domain, we propose a dual-branch nested U-shaped architecture that inherits structural information while refining image details. The outer branch maintains global structural consistency to suppress redundant generation, while the inner branch enriches fine-grained details via a compensatory attention mechanism. Extensive experiments on multiple datasets demonstrate that FacDNet achieves superior performance in both quantitative and qualitative analysis. Our source code will be publicly available at https://github.com/ZNY-Kizz/FacDNet.