Noise, Denoise, Correct: MCMC Posterior Sampling with Diffusion Priors in Three Steps
We introduce diffusion waltz, an MCMC method using SDEdit-style noising-denoising as a proposal, corrected via Metropolis-Hastings for exact posterior sampling without prior evaluation.
ProofPaper ↗
Key points
- Pretrained diffusion models are powerful priors for inverse problems, but posterior sampling under nonlinear, non-differentiable forward models remain hard.
- We further propose injecting observations into the proposal while preserving exactness, using a gradient-free ensemble Kalman update.
- On a non-differentiable Navier-Stokes initial condition recovery task, diffusion waltz outperforms existing baselines across different noise and nonlinearity regimes.
Sources (1)
- [1]Noise, Denoise, Correct: MCMC Posterior Sampling with Diffusion Priors in Three StepsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 04:06 AM
We introduce diffusion waltz, an MCMC method using SDEdit-style noising-denoising as a proposal, corrected via Metropolis-Hastings for exact posterior sampling without prior evaluation.
Pretrained diffusion models are powerful priors for inverse problems, but posterior sampling under nonlinear, non-differentiable forward models remain hard.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 7, 2026Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning
- Oct 6, 2026Learning Transition Kernels of Jump-Diffusion Processes with Conditional Diffusion Models
- Oct 6, 2026SNR-Gated LSTM-Conditioned Diffusion Model for MIMO Channel Estimation
- Oct 6, 2026Disentangling Dual Image References in Frequency Aware Diffusion Models for Personalized Generation
- Oct 6, 2026Uniform Discrete Diffusion Models are Minimax Optimal for Estimating Distributions with Small Effective Support Size
- Aug 10, 2026vllm-project/vllm v0.27.0