Prior Enhanced Diffusion Model for Low Light Image Enhancement

Low-light image enhancement is a critical challenge in computer vision, particularly in applications such as surveillance, autonomous driving, and photography. Traditional enhancement techniques often fail to preserve details and colors in poorly illuminated images. In this paper, we propose a novel diffusion models based method enhanced with effective prior for low light image enhancement. Specifically, observing the underlying structures instead of pure noise in low-light images, we first remove the strong Gaussian noise limits in each diffusion step to reserve the details. Second, to ensure the overall color consistence,a 3-channel Color Map is involved to provide essential intensity information. Our experiments reveal that removing Gaussian noise limits from model output during training significantly enhances the model’s ability to restore low-light images, demonstrating that excluding noise allows for sharper and more accurate restorations. Our findings indicate that such prior can boost the diffusion model and yield superior performance across multiple evaluation metrics. This work offers valuable insights for optimizing preprocessing techniques in diffusion-based models for low-light image enhancement, paving the way for improved real-world applications.