Research map / Probabilistic and scientific modeling / Variational and generative sampling
Moment Matching Denoising Gibbs Sampling
Mingtian Zhang, Alex Hawkins-Hooker, Brooks Paige, David Barber
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use analytical full x-tilde-dependent covariance instead of learned diagonal or isotropic covariance in Gibbs sampling (ablation) | Moment matching denoising Gibbs sampler · 2D mixture of Gaussians, Swiss roll, two rings | MMD | Improved |
| replace KL-trained learned diagonal covariance with analytic moment-matched covariance (ablation) | Gibbs sampler on MNIST · MNIST | sample quality (visual) | Improved |
| use x-tilde-dependent diagonal covariance instead of x-tilde-independent isotropic covariance (ablation) | Gibbs sampler on MNIST · MNIST | background cleanliness (visual) | Improved |
| learned diagonal covariance trained by KL minimization versus DSM objective (ablation) | MNIST energy/denoising model · MNIST | training stability | Hurt |
| vary noise level sigma among 0.3, 0.5, 0.8 (ablation) | Single-noise Gibbs sampler · MNIST | visual sample quality | Mixed |
| increase number of Rademacher samples for diagonal Hessian estimate (1, 10, 100) (ablation) | Diagonal covariance estimator · CIFAR-10 | covariance estimation quality (visual) | Improved |
| run more Gibbs steps with single noise level sigma=0.3 (ablation) | Single-noise Gibbs sampler · CIFAR-10 | FID | Hurt |
| run Gibbs chains longer on MNIST | Single-noise Gibbs sampler · MNIST | mode coverage | Hurt |
| use multiple noise levels instead of a single noise level (ablation) | Moment matching Gibbs sampler · CIFAR-10 | FID | Improved |
| replace annealed Langevin dynamics with multi-level Gibbs sampling (3 steps per level, 3 Rademacher samples) | Pretrained noise-conditioned score network · CIFAR-10 | FID and Inception score | No clear effect |
| add final denoising step returning clean-distribution mean after last noise level | Multi-level Gibbs sampler · CIFAR-10 | FID | Improved |
Closest papers by meaning
- Iterated Denoising Energy Matching for Sampling from Boltzmann Densities (ICML 2024 · similarity 0.86)
- Sampling Binary Data by Denoising through Score Functions (ICML 2025 · similarity 0.86)
- Energy-Based Modelling for Discrete and Mixed Data via Heat Equations on Structured Spaces (NeurIPS 2024 · similarity 0.86)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.85)
- Learning Energy-Based Models from Stochastic Interpolants using Spatiotemporal Differences (NeurIPS 2026 · similarity 0.85)
- Dimension-free Score Matching and Time Bootstrapping for Diffusion Models (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
