Supplementary Data for "Benchmarking Diffusion Annealing-Based Bayesian Inverse Problem Solvers”

This dataset contains approximate posterior samples from algorithms in the Bayesian Inverse Problem Solvers through Diffusion Annealing (BIPSDA) framework (journal paper: doi.org/10.1109/OJSP.2025.3597867, code: https://codeocean.com/capsule/6733743/tree). In particular, four HDF5 files corresponding to the stylized inpainting studies (low and high noise regimes), stylized x-ray tomography study, and stylized phase retrieval study are provided. Each file contains 18 datasets. These include datasets containing the ground truth values of the parameter of interest, the observed measurements, and the reference ground truth posterior samples for each of the 100 posterior trials we conducted. The remaining 15 datasets correspond to the BIPSDA algorithm runs; the naming convention is the BIPSDA algorithm name followed by an option that indicates whether the analytic or learned prior score was used in the algorithm run. In each dataset, the leading dimension corresponds to the posterior sampling trial number.

本数据集包含扩散退火贝叶斯逆问题求解器（Bayesian Inverse Problem Solvers through Diffusion Annealing，简称BIPSDA）框架下各类算法生成的近似后验样本。相关研究的期刊论文DOI为doi.org/10.1109/OJSP.2025.3597867，配套开源代码链接为https://codeocean.com/capsule/6733743/tree。具体而言，本次发布包含四个HDF5格式文件，分别对应风格化图像修复研究（涵盖低噪声、高噪声两种工况）、风格化X射线层析成像研究，以及风格化相位恢复研究。每个文件内含18个数据集，其中包含本次100次后验采样试验中，各试验对应的待估计参数真值、观测测量数据，以及参考基准后验样本数据集。剩余15个数据集对应BIPSDA算法的运行结果，其命名规则为：以BIPSDA算法名称为前缀，后接后缀以区分本次运行采用解析先验得分还是可学习先验得分。每个数据集的首个维度代表后验采样试验的序号。