Research map / Image generation / Fast diffusion sampling

Composing Diffusion Priors with Explicit Physical Context via Generative Gibbs Sampling

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Diffusion theory and learned samplers · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GG-PA, a training-free Gibbs sampler with diffusion-time replica exchange that composes partial pretrained diffusion priors with explicit physical context via an augmented joint target, tested on a double-well, a phi^4 lattice, and alanine dipeptide systems.

Problem

Pretrained diffusion priors cover only subsets of degrees of freedom, so sampling the true target under explicit physical context (solvents, ions, other subsystems) requires composing partial priors at inference time, which standard posterior-sampling methods handle poorly due to intractable margina

Approach

Formulates composition as Gibbs sampling on an augmented state space (full system state plus prior variables) indexed by diffusion time; alternates parallel denoising of each prior with a context-aware aggregation step, is asymptotically exact as t→0, exact at finite t for quadratic interactions, and uses replica exchange over diffusion time with a tractable swap rule where prior densities cancel.

Key results

• On the coupled double-well, GG-PA error is minimal at t=0.2 and grows for t above the theoretical bound t_max=0.28; at t=0.1 fixed-t GG-PA mixes slower than MD while GG-PA-RE gives the largest speedups.
• On the phi^4 lattice, GG-PA reproduces the magnetization, susceptibility peak at J_c≈0.436 and Ising scaling collapse, trained only on on-site factors.
• On alanine dipeptide dimers, GG-PA-RE qualitatively recovers topology-dependent torsion preferences but assigns more mass to the unassigned residual category than MD.

Surprising findings

• Exact finite-time recovery holds only for quadratic/linear-Gaussian interactions, and the Gaussian-split-Gibbs analysis implies a covariance correction for existing split Gibbs samplers.

Limitations

• Exact finite-t recovery relies on quadratic/linear-Gaussian structure; non-quadratic systems such as alanine dipeptide incur approximation error.
• MCMC-based sampling can mix slowly in stiff or high-dimensional systems; replica exchange requires multiple replicas.
• Performance depends on prior fidelity and decomposition quality, with risk of double-counting interactions.

Experimental setup

Datasets: MD trajectories of isolated alanine dipeptide monomer in vacuum at 300 K synthetic double-well samples synthetic phi^4 on-site samples

Benchmarks: coupled double-well phi^4 lattice model alanine dipeptide AD-Na+ and AD dimer

Metrics: Jensen-Shannon divergence integrated autocorrelation time order parameter susceptibility occupancy of torsional state categories

Baselines: Direct Diffusion molecular dynamics checkerboard Metropolis Monte Carlo

Builds on: split Gibbs sampling plug-and-play posterior sampling replica exchange Hamiltonian parallel tempering MBAR

Closest papers by meaning

• Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities (NeurIPS 2025 · similarity 0.89)
• End-to-end Learning of Gaussian Mixture Priors for Diffusion Sampler (ICLR 2025 · similarity 0.89)
• Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.88)
• On scalable and efficient training of diffusion samplers (NeurIPS 2025 · similarity 0.88)
• Generating Physical Dynamics under Priors (ICLR 2025 · similarity 0.88)
• Diffusive Gibbs Sampling (ICML 2024 · similarity 0.88)