Composing diffusion priors with explicit physical context via generative Gibbs sampling
Advances in Neural Information Processing Systems (NeurIPS), 2026. Accepted, to appear
Generative Gibbs for Physics-Aware Sampling (GG-PA). Accepted at NeurIPS 2026.
GG-PA is a training-free framework that composes pretrained diffusion priors with explicit physical context at sampling time. Learned partial priors and the physical context are treated as one joint target in an augmented space, and a Gibbs sampler alternates denoising and context-aware updates. The sampler is asymptotically exact as the diffusion time goes to zero and remains exact at finite diffusion time for quadratic interactions.
Talk: Generative Gibbs sampling: Composing generative models with explicit physical context, Microsoft Research New England, September 2026.