Mask Forcing Paper Cuts Over-Saturation in AR Video Diffusion Distillation, Wins 79-83% Human Vote

A team from HKUST(GZ), HKUST, Lightspeed, UCSD, CUHK(SZ) and NUS introduces Mask Forcing, which injects randomly masked cleaner tokens into rollout inputs to fight mode-seeking and error accumulation in autoregressive video diffusion distillation. On a 100-prompt benchmark it improves HPSv3 by up to 1.03 and instruction-following by up to 6.53 points; human raters preferred Mask Forcing over baselines 79-83% of the time, with training taking ~1.5k steps on 8 GPUs.