ForcingBooster: Adaptive Patch-wise Denoising for Video Diffusion
October 7, 2026
ForcingBooster dynamically allocates denoising steps to individual patches in autoregressive video diffusion models. It avoids redundant computation on static regions while allocating more steps to complex, high-motion patches.
HOW THIS AFFECTS YOU
●
builderYou can speed up real-time interactive video generation by optimizing compute allocation per patch.
●
researcherThis provides a framework for more efficient spatial compute budgeting in diffusion models.