Interactive figure. SnapFlow refines a one-step velocity field so a single jump from noise lands on the data. Toggle the instantaneous velocity field (curved) and the one-step field (which straightens over training), and play training to watch the one-step jump close on the data.

[ Distillation ]

One jump from noise to action

SnapFlow refines the action-expert flow field so that a single jump from noise lands on data. The instantaneous vs. one-step velocity prediction is conditioned via a 2-layer MLP on the target timestamp: the target_timestamp for the FM loss is sampled in the typical fashion from [0.001, 1.0], while the shortcut loss assigns target_timestamp = 0 to activate one-step prediction.

0%
t = 1.00