In-Flight KV Cache with Clean Anchors for Faster Autoregressive Video Diffusion
FlashForward reuses in-flight KV caches and clean anchors to speed long autoregressive video diffusion.
FlashForward speeds few-step autoregressive video diffusion by reusing the in-flight key-value cache from each denoising forward instead of extra cache-only passes. Assigning one GPU per stage lets chunks occupy different stages at once, while sparse clean anchor latents stabilize noisy history and limit appearance and motion drift. With up to four GPUs it is 1.16–1.69 times faster than HiAR and 1.42–2.92 times faster than Self-Forcing for 16 FPS videos of 20 seconds or longer, at 1.3B and 14B scales and 480p or 720p. On VBench, the 1.3B model at 480p scores higher and stays stable at 20, 35, and 65 seconds.
- FlashForward reuses KV already computed during each denoising stage.
- One GPU per stage lets different chunks run concurrently.
- Sparse clean anchors counter drift from noisy stage-matched history.
- Reported speedups cover 1.3B and 14B backbones at 480p and 720p.
From the original article
1 min read at huggingface.co
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 (KV) cache by additional forwards to build the cache without advancing an output latent.
By Yikai Wang, Xiao Han, Mengmeng Xu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.