Mask Forcing Prevents Mode Collapse in Video Diffusion Distillation
September 7, 2026
Mask Forcing introduces a Dual-Noise Masking Rollout strategy to improve autoregressive video diffusion distillation. This technique mitigates the over-saturation and over-smoothing common in DMD by preventing the student distribution from collapsing onto limited modes of the teacher.
HOW THIS AFFECTS YOU
●
builderYou can achieve higher visual realism and variety in real-time distilled video generation models.
●
researcherYou can utilize dual-noise masking to address reverse-KL mode-seeking issues in distillation.