Velocity Scaling in Flow Matching Cuts ImageNet-256 FID from 28.0 to 12.2

A new arXiv paper from Youssef Saied and François Fleuret challenges the prevailing explanation for why scaling a flow-matching velocity field improves image generation quality. Prior work attributed the gain to correcting a velocity-magnitude deficit introduced by MSE training; the authors show that no such deficit exists. Instead, they argue that velocity scaling reduces "population time lag" — a mismatch in which sampled states at model time t resemble training states from an earlier time — and that directly measuring this lag to select a scaling gain is a more principled approach.

Applied across architectures and model sizes on ImageNet-256 at 25 neural function evaluations without classifier-free guidance, the method reduces FID from 28.0 to 12.2 (estimated by linear interpolation between neighboring gain measurements). The 41-page paper, submitted October 7, 2026, also notes that shifting model time backward is an equivalent alternative to velocity scaling for addressing population time lag.