Research map / Image generation / Fast diffusion sampling
Exploring the Design Space of Transition Matching
Uriel Singer, Yaron Lipman
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 |
|---|---|---|---|
| add high-frequency stochastic sampling to D-TM MLP head (ablation) | D-TM MLP · aggregate rank | rank: 0.51 -> 0.66 | Improved |
| add low-frequency stochastic sampling to D-TM Transformer head (ablation) | D-TM Transformer · aggregate rank | rank: 0.52 -> 0.58 | Improved |
| scale head sequence (l=4) for Transformer head (ablation) | D-TM Transformer head · aggregate rank | rank: 0.43 -> 0.52 (with bs16, lognormal) | Improved |
| scale head sequence for MLP head (ablation) | D-TM MLP head · aggregate rank | rank | No clear effect |
| increase head size (ablation) | D-TM head · aggregate rank | rank | No clear effect |
| add a head (D-TM) vs dense backbone-only (ablation) | Transition Matching · aggregate rank | rank: Dense 0.25 -> DTM 0.36 | Improved |
| increase head batch size to 16 (ablation) | D-TM head · aggregate rank | rank: MLP 0.36 -> 0.51 with bs16 plus lognormal weights | Improved |
| lognormal time weighting for backbone and head (ablation) | D-TM · aggregate rank | rank: FM 0.16 -> 0.28 (FM uniform vs lognormal) | Improved |
| Y parameterization: difference X1-X0 vs denoiser X1 vs noise X0 (ablation) | D-TM head · aggregate rank | rank | Improved |
| D-TM MLP head instead of flow matching at inference | D-TM MLP · wall-clock | speedup: ~5x wall-clock speedup at higher rank (0.8 s vs 4 s) | Improved |
Closest papers by meaning
- Transition Matching: Scalable and Flexible Generative Modeling (NeurIPS 2025 · similarity 0.94)
- Equilibrium Matching: Generative Modeling with Implicit Energy-Based Models (NeurIPS 2026 · similarity 0.89)
- Generator Matching: Generative modeling with arbitrary Markov processes (ICLR 2025 · similarity 0.89)
- Inductive Moment Matching (ICML 2025 · similarity 0.89)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.88)
- One Step Diffusion via Shortcut Models (ICLR 2025 · 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.
