TAC-Diff: Texture-Anchored Color Diffusion for Low-Light Image Enhancement.

Low-light image enhancement aims to improve images captured under poor illumination by restoring texture details and correcting color and contrast. While diffusion-based models have demonstrated powerful generative capabilities and significant progress in this task, existing approaches face two empirical challenges. First, operating directly in the RGB color space, where texture information is jointly encoded across channels, may be associated with inadequate modeling of its cross-channel distribution, resulting in insufficient texture restoration. Second, the parameter optimization direction tends to place greater emphasis on reducing chrominance errors, and this color-oriented tendency may coincide with insufficient reconstruction of fine textures. Motivated by these observations, we propose TAC-Diff, an efficient Texture-Anchored Color Diffusion framework. Our method operates in a decoupled color space and integrates a diffusion model with multi-directional texture-difference awareness to refine texture details. In addition, we introduce a texture reconstruction-uncertainty representation that aids in recovering complex regions and encodes texture priors. These priors are then unidirectionally injected into the chrominance branch to extract cross-texture consistency cues, enabling robust and accurate color reconstruction. The dual-branch architecture, combined with a distribution-adapted diffusion prior, constitutes our novel TAC-Diff approach. Extensive experiments on multiple benchmark datasets demonstrate that TAC-Diff effectively preserves rich textures and natural colors with fewer parameters and lower computational complexity.