A new technique speeds up AI video generation by skipping work the model was already quietly redoing.
The system, called FlashForward, targets autoregressive video diffusion models, which build long videos one chunk at a time through several rounds of denoising. To keep later chunks consistent with earlier ones, older methods ran extra passes through the model just to rebuild a memory cache, with no new video output to show for it. FlashForward instead reuses the cache data each denoising step already produces, cutting out those wasted forward passes. Because that shortcut makes the memory noisier and lets chunks drift apart in appearance and motion, FlashForward also periodically generates sparse, clean anchor frames to keep distant chunks visually tied together.
In testing with up to four GPUs, FlashForward produced 16-fps videos of 20 seconds or longer 1.16-1.69x faster than a prior system called HiAR, and 1.42-2.92x faster than one called Self-Forcing, across 1.3B and 14B parameter models at 480p and 720p. On the VBench quality benchmark, the 1.3B model scored higher than those rivals and held up better as videos stretched to 35 and 65 seconds. That combination, speed gains without the typical quality tradeoff, is what makes longer AI-generated clips start to look computationally practical rather than theoretical.
Still, this is benchmark performance from a paper, not a deployed product. And the "clean anchors" meant to stabilize the video are themselves machine-generated guesses, not ground truth, so the method is partly solving a drift problem it also introduces.