Velocity Scaling in Flow Matching Fixes Population Time Lag, Cutting ImageNet-256 FID from 28.0 to 12.2
francoisfleuret · x · 2026-10-10
A new paper from François Fleuret's group challenges the claim that MSE-trained flow matching velocity fields underestimate velocity magnitude. Instead, velocity scaling corrects a population-level time lag in sampling. Measuring the lag to select the gain improves ImageNet-256 FID from 28.0 to 12.2 at NFE 25 without guidance, across architectures and model sizes.
More from Multimodal
- SeeDance 2.5 mostly fixes the stitching seams that plagued SeeDance 2 — TomLikesRobots · 2026-10-10
- Two Midjourney --sref style codes to try: 754853361 and 824325735 — sergeantsref · 2026-10-10
- AI mushroom-hair DJ video earns $19,433 with 9.9M views — FinanceYF5 · 2026-10-10
- Krea2 vs Qwen Image 2.1: user praises Krea2's text-to-image and feminine character rendering — morikomorizz · 2026-10-10
- Hidden 'Voice' Preview tab suggests Anthropic is building its own voice models — testingcatalog · 2026-10-10
- Creator's cloned voice narrates entire Eleven v4 demo video — aziz4ai · 2026-10-10