Apple has released FastFrames, a family of open-weight video generation models built for interactive latency rather than offline rendering. FastFrames-S, a 1.3-billion-parameter autoregressive video diffusion model, produces a frame in about 36 milliseconds on a single NVIDIA B200 GPU — roughly 28 responses per second — while the 14-billion-parameter FastFrames-L runs at 45 milliseconds per frame on an eight-GPU B200 node. Weights are on Hugging Face, code on GitHub, and the paper, "FastFrames: Video Generation at Interactive Latency," went up on arXiv as 2611.23226. The model cards are dated November 16.
The target is defined by the paper, not by Apple marketing: the authors set a bar of 50 milliseconds or less per frame — 20 or more responses per second — citing human-computer interaction research that places delay-detection thresholds around 50 to 60 milliseconds. Both models output 832x480 video in clips of 81 frames at 16 frames per second. All speed and quality figures are the authors' own measurements on B200 hardware and have not been independently reproduced.
The interesting engineering is about where the time was hiding. The team profiled state-of-the-art autoregressive video pipelines and found that steps usually treated as fixed implementation details — VAE decoding and key-value cache updates — account for over half of end-to-end inference latency. FastFrames replaces the heavy VAE decode with a lightweight pixel decoder path and treats KV-cache refresh as a first-class cost center, adding a KV refinement network and keeping generation at two diffusion steps per frame via a technique the authors call Causal Forcing++. The paper claims FastFrames-S matches the quality of a state-of-the-art model that is nearly 8 times slower, and that FastFrames-L is comparable to Krea-Realtime on realism, prompt adherence, composition and text rendering while running more than 10 times faster.
Alongside the base models, Apple released an optimized version of LingBot 1.3B that it says is 16 times faster than the official implementation and supports camera-conditioned control — steering the view with a keyboard while the video generates, in a browser demo. That is the interaction model FastFrames is designed for: live camera steering and physics-driven changes while frames keep coming, closer to a game engine loop than a text-to-video batch job.
The provenance chain is unusually complete. FastFrames-L was distilled from Wan2.1-14B through a four-checkpoint ladder — causal teacher, consistency-distilled two-step model, DMD-distilled model, then the final generator plus pixel decoder and KV refinement networks — and Apple has published every checkpoint in that ladder, so other teams can restart the distillation from any stage. The work is credited to Gene Chou, listed with Apple and Cornell University, together with Apple co-authors including Guandao Yang, Fujun Luan, Lars Mescheder and Vladlen Koltun; the paper says it was completed during Chou's internship at Apple.
The license terms deserve a careful read. FastFrames ships under the Apple Sample Code License, and parts derived from LingBot are under Creative Commons Attribution-NonCommercial-ShareAlike 4.0. Researchers can inspect and run everything, but this is not permission to embed the weights in a commercial product, and nothing in the release suggests a consumer feature is imminent — this is a research drop, not an iPhone announcement.
Still, the release marks a shift in what Apple is willing to put into the open. An interactive-latency video model is the missing piece for embodied agents that need to watch the world and react, for game prototyping, and for camera-steerable generation on-device. By publishing the latency profile itself — showing that half the wait lived outside the diffusion step — Apple has handed every competing video team a checklist, whether or not they ever touch the weights.
Comments (0)
Log in to join the discussion
Log InNo comments yet