Efficient Bidirectional Diffusion Autoregressive Video Synthesis
October 13, 2026
Glance Forcing reduces the prohibitive cost of distilling bidirectional diffusion models into causal student models by using a single-sample trajectory supervision setting. This addresses the high GPU-hour and storage requirements typically needed for offline PF-ODE trajectories.
HOW THIS AFFECTS YOU
●
builderYou can potentially train causal video world models with significantly lower computational overhead.
●
researcherThe method investigates the relationship between trajectory geometry and quality collapse during single-sample supervision.