Research map / Image generation / Fast diffusion sampling
Likelihood Matching for Diffusion Models
Lei Qian, Wu Su, Yanqi Huang, Song Xi Chen
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 Likelihood Matching with diagonal Hessian (r=0) instead of score matching | diffusion model on CIFAR-10 · CIFAR-10 | FID: 3.15 -> 3.12 | Improved |
| increase Hessian rank to r=20-30 (ablation) | Likelihood Matching · CIFAR-10 | FID: 3.12 -> 3.01 | Improved |
| increase Hessian rank to r=100 or 200 (ablation) | Likelihood Matching · CIFAR-10 | FID: 3.01 -> 3.09 | Hurt |
| use Likelihood Matching (r=30) instead of score matching | diffusion model on CelebA · CelebA 64x64 | FID: 2.71 -> 2.62 | Improved |
| use Likelihood Matching (r=30) instead of score matching | diffusion model on CIFAR-10 · CIFAR-10 | NLL (bpd): 3.28 -> 3.11 | Improved |
| learn the Hessian rather than fix it to identity (ablation) | Likelihood Matching · MNIST (N=2) | FID: 10.28 (identity Hessian) vs full model lower | Improved |
| increase number of transition steps N and Hessian rank r (ablation) | Likelihood Matching · MNIST | FID | Improved |
| use Likelihood Matching instead of score matching on Gaussian and t mixtures | MLP score and Hessian models · 1D mixture (MMD) | MMD | Improved |
| use Likelihood Matching (N=2) instead of score matching | Swiss-roll distribution · 2D Swiss-roll (log-MMD) | log-MMD | Improved |
| use Likelihood Matching with Hessian-based sampler at few reverse steps | MNIST sampler · MNIST | FID | Improved |
| use Likelihood Matching instead of score matching on ImageNet 224x224 | training cost · ImageNet 224x224 | training time and memory: 3-4x training time, 2-3x memory | Hurt |
Closest papers by meaning
- Improving Probabilistic Diffusion Models With Optimal Diagonal Covariance Matching (ICLR 2025 · similarity 0.89)
- O(d/T) Convergence Theory for Diffusion Probabilistic Models under Minimal Assumptions (ICLR 2025 · similarity 0.89)
- Dimension-free Score Matching and Time Bootstrapping for Diffusion Models (NeurIPS 2025 · similarity 0.89)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.89)
- Diffusion Models With Learned Adaptive Noise (NeurIPS 2024 · similarity 0.89)
- Reverse Diffusion Monte Carlo (ICLR 2024 · similarity 0.88)
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.
