In-Flight KV Cache with Clean Anchors for Faster Autoregressive Video Diffusion

A new technique called In-Flight KV Cache with Clean Anchors speeds up few-step autoregressive video diffusion by caching key-value pairs from already-generated temporal chunks, according to the research. Previous methods reconstruct a clean or less-noisy key-value cache to memorize generated chunks, and the new approach instead maintains the cache in flight using clean anchors. The method targets the chunk-by-chunk denoising pipeline that autoregressive video diffusion models use to generate long videos.

Few-step autoregressive video diffusion generates a long video by splitting the video into temporal chunks and generating chunk-by-chunk, each through a short sequence of denoising stages. To memorize chunks that are already generated, previous methods reconstruct a clean or less-noisy key--value K