Disen-Forcing: Disentangling Semantic Anchoring from Motion for Autoregressive Video Diffusion
A small subset of attention heads keeps generated video consistent over long sequences; most heads handle local motion.
The authors find that giving every head persistent access to initial frames can suppress motion during distillation. Their Disen-Forcing plug-in separates how heads receive that context and adds regularization to reinforce their distinct roles. On long-horizon benchmarks, they report improvements in both consistency and motion dynamics over the base models, a balance existing frame-sink methods fail to achieve.
HF Daily Papers' note
The authors find that giving every head persistent access to initial frames can suppress motion during distillation. Their Disen-Forcing plug-in separates how heads receive that context and adds regularization to reinforce their distinct roles. On long-horizon benchmarks, they report improvements in both consistency and motion dynamics over the base models, a balance existing frame-sink methods fail to achieve.
HF Daily Papers' note
score 5