In-Flight KV Cache with Clean Anchors for Faster Autoregressive Video Diffusion
Yikai Wang, Xiao Han, Mengmeng Xu, Juan Camilo Perez, Yiannis Douratsos, Sen He, Zijian Zhou, Fei Zhang and 4 others
Meta · Nanyang Technological University
Published September 26, 2026 · on the feed September 29, 2026 · ▲ 33 · arXiv 2609.32540
The abstract
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. However, every denoising forward itself already computes the in-flight KV of the current chunk. We introduce FlashForward, which directly reuses this cache to avoid the heavy cache-update-only model forwards. After the current chunk completes one denoising stage, its stage-specific cache is already available for the next chunk. Assigning one GPU to each stage therefore lets different chunks occupy different stages concurrently. This early availability has a quality cost: the resulting stage-matched history is noisy, causing appearance and motion drift among chunks. To complement it, FlashForward produces sparse auxiliary clean anchor latents before the corresponding region is generated so the generation trajectories can be stabilized by this two-sided conditioning. The two memories operate at different temporal scales: sparse clean anchor KV supplies coarse, long-range two-sided structural guidance, while dense stage-matched history preserves fine, recent evolution. With up to four GPUs, FlashForward runs 1.16--1.69times faster than HiAR and 1.42--2.92times faster than Self-Forcing for 16 FPS videos of 20 seconds or longer across 1.3B and 14B backbone scales at 480p and 720p. On VBench, for the 1.3B model at 480p, it achieves higher scores and remains stable at longer durations, demonstrating that FlashForward generates high-quality and temporally consistent videos across durations of 20s, 35s and 65s at a much faster generation speed.
The Brief
2:52A short podcast summary of the paper. Written and voiced by AI.
Transcript
In-Flight KV Cache with Clean Anchors for Faster Autoregressive Video Diffusion. By Yikai Wang and eleven others, from Meta and Nanyang Technological University. Published on the twenty-sixth of September. No code repository is linked to it.
Few-step autoregressive video diffusion makes a long video by cutting it into temporal chunks and generating them one after another, each through a short sequence of denoising stages. The model has to remember the chunks it has already made. Previous methods do this by building a clean, or at least less noisy, key-value cache, and building it costs extra forward passes that exist only to update the cache and do not advance any output. Yet every denoising forward pass already computes the key-value entries of the current chunk as a matter of course. That in-flight cache is simply not being used.
FlashForward reuses it directly and drops the heavy forward passes that only update the cache. Once the current chunk finishes one denoising stage, the cache for that stage is already available to the next chunk. That early availability opens up parallelism. Give each stage its own GPU, and different chunks can occupy different stages at the same time, like items moving along a production line.
The early cache has a cost in quality. The history it provides is matched to a stage, which means it is still noisy, and that noise causes appearance and motion to drift from chunk to chunk. To counter it, FlashForward produces sparse auxiliary clean anchor latents ahead of time, before the region they correspond to is generated. Generation is then conditioned from both sides, and that two-sided conditioning stabilizes the trajectories.
The two memories work at different time scales. The sparse clean anchors give coarse, long-range structural guidance from both directions. The dense stage-matched history keeps the fine detail of what happened most recently. One holds the overall shape of the video steady, and the other keeps its motion continuous.
With up to four GPUs, FlashForward runs between one point one six and one point six nine times faster than HiAR, and between one point four two and two point nine two times faster than Self-Forcing. Those figures are for sixteen frames-per-second video lasting twenty seconds or more, at resolutions of four hundred eighty p and seven hundred twenty p, across backbone scales of one point three billion and fourteen billion parameters.
Quality is checked on VBench with the one point three billion parameter model at four hundred eighty p. There, FlashForward scores higher and stays stable as durations grow. The authors report high-quality, temporally consistent videos at durations of twenty, thirty-five and sixty-five seconds, generated at a much faster speed.
Previous methods paid for memory with extra computation. FlashForward keeps the memory that denoising was already producing, accepts that it arrives noisy, and spends a small number of clean anchors to hold the video steady. The speed comes from the reuse and the pipelining, and the anchors are what keep that speed from costing coherence.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Long videos are made one chunk at a time. The wall here is the time spent remembering.
You are generating a long video one short chunk at a time. Each chunk is cleaned up through a few denoising stages, and later chunks need a memory of the earlier ones.
The road ahead
- 1 · The problem ~6 min Long videos are made chunk by chunk, and the extra passes spent only on building memory of past chunks now slow everything down.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Use the keys and values each denoising pass already computes as the next chunk's memory, run stages in parallel, and add clean anchors against drift.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min A planner lays down clean anchors first; then each ordinary denoising pass advances a chunk and hands its cache to the next chunk.
- 6 · The experiment ~6 min You’ll guess: With up to four GPUs, at best how much faster than Self-Forcing did FlashForward generate long videos by reusing the in-flight cache?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.