Research map / Probabilistic and scientific modeling / Variational and generative sampling
Terminal Velocity Matching
Linqi Zhou, Mathias Parger, Ayaan Haque, Jiaming Song
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 |
|---|---|---|---|
| Match terminal velocity (TVM) instead of MeanFlow initial-velocity matching | DiT-XL/2 one-step generator · ImageNet-256 1-NFE | FID: 3.43 -> 3.29 | Improved |
| Use TVM with 4 NFEs instead of 500-NFE DiT | ImageNet-256 · ImageNet-256 | FID: 2.27 -> 1.99 | Improved |
| Use TVM with 4 NFEs vs DiT on ImageNet-512 | ImageNet-512 · ImageNet-512 | FID: 3.04 -> 2.94 | Improved |
| Sample constant CFG instead of random CFG during training (ablation) | TVM · ImageNet-256 1-NFE, w=2 | FID: 5.14 -> 4.81 | Improved |
| Use EMA target (gamma=0.99) instead of no EMA (ablation) | TVM pseudo-target · ImageNet-256 1-NFE | FID: 10.24 -> 4.90 | Improved |
| Add scaled parameterization with CFG weight (w=1.5) (ablation) | TVM · ImageNet-256 1-NFE | FID: 9.32 -> 6.04 | Improved |
| Add scaled parameterization (w=2) (ablation) | TVM · ImageNet-256 | FID 1-NFE / 2-NFE: 1-NFE 3.82 -> 3.72; 2-NFE 3.27 -> 3.35 | Mixed |
| Train with t=s (pure FM) samples 10-30% of the time (ablation) | TVM · ImageNet-256 | FID 1-NFE / 2-NFE: 1-NFE 3.72 -> 3.91; 2-NFE 3.35 -> 2.97 (20%) | Mixed |
| Sample t biased toward 1 and s toward 0 (truncated sampling) (ablation) | TVM time sampling · ImageNet-256 1-NFE | FID: 4.59 -> 3.70 | Improved |
| Over-bias time sampling (mu_t=2.0, sigma_t=2.0) (ablation) | TVM time sampling · ImageNet-256 1-NFE | FID: 3.70 -> 7.88 (varied configs) | Hurt |
| Use gap sampling with lower mean gap (ablation) | TVM time sampling · ImageNet-256 1-NFE | FID: 5.12 -> 3.72 | Improved |
| Use higher CFG weight in training (ablation) | TVM · ImageNet-256 | FID 1-NFE vs 2-NFE | Mixed |
Closest papers by meaning
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.90)
- Reflected Flow Matching (ICML 2024 · similarity 0.90)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.90)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.89)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.88)
- Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling (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.
