Radar / AI media tools / FastFrames
FastFrames: Apple's open video models that generate frames in 36 ms
Apple and Cornell's family of autoregressive video diffusion models that generate 480p frames at interactive latency — the 1.3B variant runs at 36 ms per frame on a single B200 GPU while matching the quality of a model nearly 8x slower.
Why it matters
Replaces the VAE decoder and KV-cache updates — the hidden half of video-generation latency — with lightweight alternatives, making live, steerable video generation practical.
What you could build with it
A game studio could build a live previs tool where a director drags objects or steers the camera and the scene re-renders in real time, replacing overnight render queues for early-stage shot planning.
Does it hold up?
Too early to judge: the weights and code landed Nov 17, the licence is research-only, and the headline speeds need B200-class hardware — no independent reproduction has appeared yet.
Built with FastFrames
- FastFrames: Video Generation at Interactive Latencyarxiv.org · The full paper: 36 ms per frame for the 1.3B model and 45 ms for the 14B model, with VBench scores matching much slower baselines.
- Apple-Cornell FastFrames Pushes AR Video to 36ms/Frame on B200aiweekly.co · Plain-English breakdown of the latency profiling and the ConvNeXt pixel head that replaces the VAE decoder.
- Apple FastFrames: open video models at interactive speedthereelstack.com · Hands-on notes: weights on Hugging Face, code public Nov 17, research-only licence, and what the 480p limit means in practice.
Learn more
First spotted on hf: source.
More AI media tools
Hyperframes: write HTML, render video — HeyGen's open video engine for AI agentsOpen-source renderer from HeyGen that turns HTML, CSS and JavaScript into MP4 video, built so AI coding agents can…media · JEV 0.78HeyGen Avatar 6: an audio-driven avatar model with script-aware gesturesHeyGen's next-generation audio-driven personal avatar model — script-aware gestures, native 2K output, a personal…media · JEV 0.52Kling Image 4.0: native 8K images with 2,000 characters of clean textKuaishou's Kling Image 4.0 generates natively in 8K, renders up to 2,000 characters of legible text in one image, and…media · JEV 0.48Suno Visuals: AI music videos and album art for your songsSuno's new Visuals feature turns songs into lyric videos with lip-sync, album art, and clips up to 15 seconds that…media · JEV 0.46MAI Image 2.7: Microsoft's image editor tops the leaderboard at a fifth of the priceMicrosoft's MAI Image 2.7 leads Artificial Analysis's image-editing leaderboard at 1,215 Elo — ahead of GPT Image 2.6…media · JEV 0.46Seedance 2.5 Flash: ByteDance's video model at twice the speed, a third less costA lighter Seedance 2.5 that generates 4-30s video clips with audio at 480p/720p, about twice as fast and roughly 34%…media · JEV 0.43
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.