New Diffusion Posterior Sampling Method Improves Inverse Problem Solutions

A recent paper on arXiv (2608.15144) presents a novel approach for posterior sampling in diffusion inverse problems, tackling the challenge posed by the conditional score's complexity. The researchers introduce a one-parameter family of posterior SDEs, where a stochasticity parameter governs both probability-flow transport and stochastic exploration while maintaining the integrity of posterior marginals. To enhance model feasibility, they reformulate the likelihood in rescaled clean-image coordinates and utilize log-SNR for structuring posterior proxies. By projecting diffusion uncertainty via the forward operator, they derive a noise-conditioned covariance path that converges towards the clean posterior. To ensure effective transport, they integrate a frozen-target Langevin corrector with transport, resulting in a continuous surrogate SDE, which is discretized through an outer Lie–Trotter splitting and variance-reduction method. This paper is classified as a cross-type announcement and can be accessed on arXiv.