Apple open-sources FastFrames: two-step diffusion for interactive-latency 480p video
_akhaliq · x · 2026-11-18
Apple released FastFrames on Hugging Face, video generation models targeting interactive latency at 480p.
- Two sizes: FastFrames-S (1.3B) and FastFrames-L (14B), both autoregressive two-step video diffusion models
- Latency: FastFrames-S on one GPU with full KV cache runs 35.4ms per step (paper: 36ms), DiT forward 16ms
- Also ships an optimized LingBot 1.3B that is 16x faster than the official model with fluid camera-conditioned control
- Repo includes latency benchmarks, an interactive demo, and T2V scripts; requires Linux + NVIDIA GPU + CUDA 12.8
Interactive-latency generation is a notable signal for real-time video apps and world-model work.
Related event: Apple Open-Sources FastFrames for Real-Time Video Generation(4 posts)→
More from Embodied
- Starkey's AI Hearing Aids Top $1B in Sales: 'The Ear Is the New Wrist' — BrandonSawalich · 2026-11-18
- binarybits: Tesla Robotaxi faces the same city-by-city slog as Waymo — binarybits · 2026-11-18
- NVIDIA GEAR researcher to present GR00T 2 Turbo world action models for humanoids at Penn GRASP — zhengyiluo · 2026-11-18
- California bans Waymo autonomous vehicles from state freeways — armcat · 2026-11-18
- Teleoperated humanoid robot TAU cooks a burger in a real kitchen — Distinct-Question-16 · 2026-11-18
- Ropedia Xperience 2: 1,500+ real scenes for scaling physical AI training data — _akhaliq · 2026-11-18