The authors first analyze the population-level mean-squared-error objective and show that its optimal velocity field transports the source distribution to the data distribution exactly.
Velocity scaling improves flow models by correcting sampling time lag
The authors show that coarse sampling leaves generated states behind their assigned model time, and that scaling the velocity field largely corrects this mismatch.
Big Tech
Youssef Saied · François Fleuret
University of Geneva · Meta FAIR
Research Digest··2 min read
Saied and Fleuret challenge the explanation that mean-squared-error training systematically makes flow-matching velocities too small.
Why this paper
From Meta FAIR and University of Geneva
In one line
Velocity scaling improves flow matching by reducing population time lag, not by correcting an MSE velocity deficit.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§