Velocity Scaling
An adjustment technique applied during flow matching inference that multiplies the trained velocity field by a time-varying gain coefficient to improve generation quality. It addresses the time-alignment problem in sampling dynamics by compensating for discrepancies between training and inference trajectories.
Timeline (last 90 days)
A study published on arXiv (arXiv:2610.10823v1) offers a different answer from the mainstream explanation regarding why velocity scaling in flow matching is effective.
The previously dominant view held that velocity fields trained with mean squared error (MSE) systematically underestimate velocity magnitudes, and that velocity scaling is a correction for this error.
The study proposes that the true reason velocity scaling is effective is population time lag.
Velocity scaling and shifting the model time backward are essentially two equivalent approaches to addressing the population time lag problem.
Directly measuring population time lag and using it to select the gain coefficient γ(t) can systematically improve generation quality without repeated trial and error.
All Facts (5)
A study published on arXiv (arXiv:2610.10823v1) offers a different answer from the mainstream explanation regarding why velocity scaling in flow matching is effective.
50%UnverifiedThe previously dominant view held that velocity fields trained with mean squared error (MSE) systematically underestimate velocity magnitudes, and that velocity scaling is a correction for this error.
50%UnverifiedThe study proposes that the true reason velocity scaling is effective is population time lag.
50%UnverifiedVelocity scaling and shifting the model time backward are essentially two equivalent approaches to addressing the population time lag problem.
50%UnverifiedDirectly measuring population time lag and using it to select the gain coefficient γ(t) can systematically improve generation quality without repeated trial and error.
50%