Apple open-sources FastFrames: 480p video generation at ~36ms interactive latency
victormustar · x · 2026-11-18
Apple released the official implementation of FastFrames on Hugging Face, targeting video generation at interactive latency:
- Two autoregressive, two-step video diffusion models: FastFrames-S (1.3B) and FastFrames-L (14B), generating 480p video at interactive latency;
- FastFrames-S measures 35.4ms end-to-end on a single GPU at 480p with full KV cache (paper: 36ms);
- Also ships an optimized LingBot 1.3B that is 16x faster than the official model with fluid camera-conditioned control and a local interactive demo;
- Repo includes latency benchmarks, T2V scripts and training code; requires NVIDIA GPU + CUDA 12.8.
Related event: Apple Open-Sources FastFrames for Real-Time Video Generation(3 posts)→
More from Multimodal
- Redditor: Grok's basic plan output looks better than the $250M movie The Flash — Ok_Item9755 · 2026-11-18
- Microsoft releases MAI-Image-2.7 Flash: cheaper, faster image generation on OpenRouter — AmyKateNicho · 2026-11-18
- Creator builds an all-AI beauty podcast with HeyGen Avatar 6 in about an hour — toolstelegraph · 2026-11-18
- AI video: Robert Pattinson's Batman shakes hands with his own James Bond — Beginning_Gur7652 · 2026-11-18
- Magnific launches Illusion: restyle any video with Seedance 2.5, available via MCP — charis_ai · 2026-11-18
- AI-generated short film 'We're Family' Episode 1: four dogs, one Sydney night — Outrageous-Map-3228 · 2026-11-18