THE MILAD STASH

Apple's FastFrames makes AI video fly, with frames in under 50 milliseconds

SUMMARY

Apple and Cornell researchers introduced FastFrames, a family of autoregressive video diffusion models that generate 480p video at interactive latency, under 50 milliseconds. They found that VAE decoding and KV cache updates account for over half of inference latency, and they replaced those parts with lightweight alternatives. The 1.3B model runs at 36 ms per frame on one B200 GPU, matching a model nearly 8× slower. The 14B model runs at 45 ms per frame on eight B200s. Optimizing LingBot with FastFrames gave a 16× speedup.

Why it matters: Generating video at under 50 milliseconds per frame brings AI video closer to real-time, interactive use, and shows that fixing decoding and caching bottlenecks, not just bigger models, can sharply cut latency.

apple-aiml-research.github.io · Nov 21, 2026, 11:55 AM PT
Read it at apple-aiml-research.github.io

Comments

No comments yet.

Comments show once Milad approves them.

More on AI

Back to the front page