BriefGPT.xyz
Mar, 2025
FlowDPS:基于流的逆问题后验采样
FlowDPS: Flow-Driven Posterior Sampling for Inverse Problems
HTML
PDF
Jeongsol Kim, Bryan Sangwoo Kim, Jong Chul Ye
TL;DR
本研究解决了在流模型框架下逆问题求解的空白,通过将扩散逆求解器扩展到流模型,提出了一种新方法FlowDPS。研究表明,FlowDPS在四个线性逆问题中的表现优于当前最先进的替代方案,且无需额外训练。
Abstract
Flow Matching
is a recent state-of-the-art framework for
Generative Modeling
based on ordinary differential equations (ODEs). While closely related to diffusion models, it provides a more general perspective on <
→