Ratio-Residual Diffusion Model for Image Restoration

Zizheng Yang, Jie Huang, Hu Yu, Man Zhou, Bing Li, Feng Zhao

Keywords: Diffusion Model, Image Restoration

Abstract: Most existing diffusion-based image restoration methods suffer from poor interpretability and inefficient sampling, due to their direct incorporation of degraded images as conditions within the original diffusion models. Recently, some researches have tried to build a new diffusion model by transferring the discrepancies between degraded and clear images, however, they cannot effectively model diverse degradation. To address these issues, we propose a universal diffusion model for image restoration that can cover different types of degradation. Specifically, our method consists of a Markov chain that convert a high-quality image to its low-quality counterpart. The transition kernel of this Markov chain is constructed through the ratio and residual between the high-quality and low-quality images, which provides a general expression that can effectively handle various degradation processes. Moreover, we analyze the characteristics of different degradation, and design a mean schedule that enables flexible control over the diffusion speed pertaining to different degradation, which yields better restoration performance. Extensive experiments have demonstrate that our method surpasses existing image restoration methods and achieves superior performance on multiple image restoration tasks, including deraining, dehazing, denoising, deblurring and low-light enhancement.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1805