Direct Intermediate Initialization for Tilted Diffusion Samplers

Forum topic · 小凯 · 2026-10-07

Summary

This paper (arXiv:2610.06834, October 2026) by Gregory D. Bellchambers introduces a direct intermediate initialization technique for diffusion posterior samplers. The key observation is that Gaussian-tilted intermediate targets along the reverse diffusion process can be pulled back to less-conditioned clean-space posteriors, with samples transported analytically to the corresponding noisy-space targets via Gaussian bridges. For the sequential Monte Carlo (SMC) sampler MCGdiff, the pulled-back problem's effective observation variance can be up to twice the diffusion noise variance. Leveraging this, an approximate solver samples a softened clean-space posterior, a Gaussian bridge maps samples to the tilted target, and only the remaining SMC suffix is run—trading asymptotic consistency for finite-particle performance. Using moment matching posterior sampling (MMPS) as solver, the hybrid method improves sliced Wasserstein distance by roughly 2x on structured Gaussian mixture inverse problems, and by more than an order of magnitude when posterior correlation modes are rare under the prior.

Paper Overview

Field: Machine Learning Author: Gregory D. Bellchambers Published: 2026-10-05 arXiv: 2610.06834

Key Idea

Some diffusion posterior samplers construct Gaussian-tilted intermediate distributions along the reverse process. The paper observes that these targets can be pulled back to less-conditioned clean-space posteriors, and samples can be transported analytically to the corresponding noisy-space targets through Gaussian bridges. For the sequential Monte Carlo (SMC) sampler MCGdiff, the effective observation variance of the pulled-back problem can be up to twice the diffusion noise variance.

Method

The authors exploit this structure to directly initialize MCGdiff at intermediate times:

1. An approximate solver samples a softened clean-space posterior. 2. A Gaussian bridge analytically maps these samples to the tilted target. 3. Only the remaining SMC suffix is run.

This trades asymptotic consistency for finite-particle performance.

Results

Using moment matching posterior sampling (MMPS) as the solver, the hybrid method improves sliced Wasserstein distance by roughly 2x on structured Gaussian mixture inverse problems.
Improvement exceeds an order of magnitude when posterior correlation modes are rare under the prior.
A control experiment that keeps the bridge but removes clean-space conditioning (prior initialization) shows most of the improvement on the standard MCGdiff Gaussian mixture benchmark is insensitive to conditioning.
Conditional initialization yields further consistent gains on structured problems, and becomes decisive on rare-mode problems—resampling cannot repopulate modes absent from the initial particle population.

Original Abstract

We observe that Gaussian-tilted intermediate targets in diffusion posterior samplers can be pulled back to clean-space posteriors and transported via a Gaussian bridge. This enables direct intermediate initialization for MCGDiff, trading asymptotic consistency for finite-particle performance with roughly 2x improvement in sliced Wasserstein distance.