Speculative Forcing for Efficient Autoregressive Video Generation

Researchers introduced Speculative Forcing, a framework for efficient autoregressive video generation that predicts when individual tokens can be accepted early without further denoising. The method treats intermediate clean estimates as drafts and freezes accepted tokens as final outputs while retaining their cached keys and values for context. Across short-video, long-video, and world generation tasks, Speculative Forcing preserves full-step sampling quality while substantially reducing computation.