Title: Generative diffusion posterior sampling for informative likelihoods Show Chinese title

Title: 用于信息似然的生成扩散后验采样

Abstract: Sequential Monte Carlo (SMC) methods have recently shown successful results for conditional sampling of generative diffusion models. In this paper we propose a new diffusion posterior SMC sampler achieving improved statistical efficiencies, particularly under outlier conditions or highly informative likelihoods. The key idea is to construct an observation path that correlates with the diffusion model and to design the sampler to leverage this correlation for more efficient sampling. Empirical results conclude the efficiency. Show Chinese abstract

Abstract: 顺序蒙特卡罗（SMC）方法最近在生成扩散模型的条件采样中表现出成功的结果。 在本文中，我们提出了一种新的扩散后验SMC采样器，在异常值条件或高度信息丰富的似然情况下实现了改进的统计效率。 核心思想是构建一个与扩散模型相关的观测路径，并设计采样器以利用这种相关性进行更高效的采样。 实证结果证明了效率。