TRACK: Accelerating Video Diffusion via Heterogeneous Model Routing
September 25, 2026
TRACK reduces video diffusion inference costs by switching between large and small models during denoising. It uses a top-K disagreement score calibrated against a reference trajectory to determine when to route tasks to less computationally expensive models.
HOW THIS AFFECTS YOU
●
builderYou can reduce inference latency and compute costs for video generation by routing steps between model sizes.