Diffusion models are increasingly used as priors in Bayesian inverse problems, but many existing posterior sampling methods depend on heuristic approximations to the generative process. The authors propose an ensemble-based algorithm that performs posterior sampling without such approximations, building on earlier work that combined diffusion models with sequential Monte Carlo.

By tracking how the prior evolves through the diffusion process encoded by a pre-trained score function, they derive a modified partial differential equation for the posterior distribution. This PDE contains a modified diffusion term and a reweighting term, and can be simulated with stochastic weighted particle methods. Theoretically, they prove that the error between the true posterior and the empirical distribution of generated samples is bounded by the training error of the score function and the number of particles.

Empirically, the method is validated on several imaging inverse problems, where it yields more accurate reconstructions than existing diffusion-model-based approaches. The paper, accepted by TMLR 2026, includes code on GitHub.