The paper presents a novel approach to enhancing low-light images using a diffusion-based framework. This method uses a regularization process on the inherent ODE-trajectory of diffusion models, inspired by the stability and effectiveness of low curvature ODE-trajectories. The technique uses global structure-aware regularization to maintain complex details and improve contrast during the diffusion process. This approach reduces noise and artifacts that may occur during the diffusion process, leading to more accurate and flexible enhancements. An uncertainty-guided regularization technique is also introduced to improve learning in challenging regions of the image. The proposed framework, combined with rank-informed regularization, demonstrates significant performance in low-light enhancement, improving image quality, suppressing noise, and amplifying contrast. The method could further stimulate exploration and improvement in low-light image processing and other diffusion model applications.
Publication date: 26 Oct 2023
Project Page: https://github.com/jinnh/GSAD
Paper: https://arxiv.org/pdf/2310.17577
