TRD Method Enables One-Step Generation via Trajectory Regularization
November 11, 2026
Trajectory-regularized distribution distillation (TRD) combines distribution-level and trajectory-level supervision using a momentum-mixture policy in a single forward pass. This removes the need for a second trainable network when estimating reverse Kullback-Leibler gradients for flow matching models.
HOW THIS AFFECTS YOU
●
builderThis could lead to faster, single-step image or video generation models without extra training overhead.
●
researcherYou can simplify one-step distillation by using a single student pass to estimate marginal velocity.