Speculative Forcing for Efficient Autoregressive Video Generation
Few-step autoregressive video generators enable streaming generation, yet they run every token through the final denoising step, regardless of how much refinement it needs. We find this uniform computation is often unnecessary: many tokens already closely match their eventual full-step outputs in the initial steps, while others require refinement until the end. We introduce Spe
Reading in Plain English mode — toggle below to switch
Few-step autoregressive video generators enable streaming generation, yet they run every token through the final denoising step, regardless of how much refinement it needs. We find this uniform computation is often unnecessary: many tokens already closely match their eventual full-step outputs in the initial steps, while others require refinement until the end. We introduce Spe
This research introduces a novel inference optimization technique, Speculative Forcing, specifically for autoregressive video generation. It addresses the inefficiency of uniform computation in denoising steps by selectively freezing tokens, leading to significant cost savings and faster generation.