Low-Light Image Enhancement via Diffusion Models with Semantic Priors of Any Region

With the emergence of the diffusion model, its powerful regression capabilities have significantly boosted the performance for low-light image enhancement. However, the inherent information loss in low-light conditions calls for a deep understanding of scene semantics and structures to effectively recover missing content. Recent advances such as the Segment Anything Model (SAM) provide semantic priors for arbitrary regions through prompt-based object segmentation, which offers rich contextual cues to guide the restoration process. Motivated by this, we propose to incorporate such semantics-aware priors into a generative diffusion framework from three perspectives. Firstly, we propose a novel Context-Aware Understanding Guided Diffusion model (CUGD) for low-light image enhancement. This method utilizes the diffusion technique to model the distribution of images by incorporating contextually aware semantic and structural information for any region. Specifically, regional priors provided by SAM are integrated to guide the diffusion process with awareness of any object or region, enhancing the model’s capability to reason about scene content. Secondly, we design a Context Understanding Injection Encoder (CUIE) module that combines self-attention and cross-attention mechanisms to comprehensively integrate semantic and structural information into enhanced results, thus facilitating a fine-grained understanding and enhancement process. This module serves the diffusion model in generating normal-light images with richer and more semantically consistent details. Lastly, the semantic context regularization loss is introduced into the optimization process, ensuring that the recovered context better aligns with the normal-light semantic distribution. Extensive experiments on various datasets show that the proposed method attains state-of-the-art (SOTA) performance in both full-reference and no-reference evaluation measures.