Stage-wise Distortion-Perception Traversal in Zero-shot Inverse Problems with Diffusion Models
Jiawei Zhang, Ziyuan Liu, Leon Yan, Zhenyu Xiao, Yuantao Gu
Read on arXiv →Key claim
MAP-RPS enables effective D-P traversal in inverse problems.
This paper presents a new framework called MAP-RPS for navigating the distortion-perception tradeoff in diffusion models. It combines MAP estimation with posterior sampling to improve perceptual quality in inverse problems. The results show that this approach effectively enhances performance across various tasks.
In plain English
This paper presents a new framework called MAP-RPS for navigating the distortion-perception tradeoff in diffusion models. It combines MAP estimation with posterior sampling to improve perceptual quality in inverse problems. The results show that this approach effectively enhances performance across various tasks.
The proposed MAP-RPS framework introduces a significant new method for D-P traversal in diffusion models, extending their applicability.
The paper provides theoretical analyses and extensive experiments demonstrating the effectiveness of the proposed methods.