FastFrames-L / FastFrames-S
Intro
FastFrames-L / FastFrames-S is the family of autoregressive video diffusion models Apple introduced to generate 480p video at interactive latency (officially <50 ms). The official project page is titled “FastFrames: Video Generation at Interactive Latency,” and the TL;DR says FastFrames-S (1.3B) and FastFrames-L (14B) “are autoregressive video diffusion models that generate 480p video at interactive latency (<50 ms).” Authors are Gene Chou (Apple, Cornell) and Apple’s Guandao Yang, Fujun Luan, Lars Mescheder, Wei Dong, Stephan R. Richter, and Vladlen Koltun.
Reported latency is 36 ms per frame, or about 28 RPS, for S on a single NVIDIA B200, matching the quality of a state-of-the-art model that is nearly 8× slower; and 45 ms per frame, or about 22 RPS, for L on an 8×B200 node, with better visual quality and realism. The same page’s comparison table gives end-to-end steady-state figures at 480p: FastFrames-S at 36 ms (1 GPU) / 14 ms (8 GPUs), and FastFrames-L at 158 ms (1 GPU) / 45 ms (8 GPUs).
Official weights are on Hugging Face at apple/FastFrames-L and apple/FastFrames-S. This scan saw FastFrames-L created at 2026-11-16T17:21:04Z; the project page and later public coverage present the family as open-weight models that can run at interactive latency. Do not confuse this with a consumer Apple-device feature: the reported timings are on NVIDIA B200 GPUs.
Advertisement
Explain
The official problem statement is that existing autoregressive video diffusion models are still at least an order of magnitude too slow for fluid interaction. FastFrames profiles the inference pipeline and argues that operations often treated as fixed implementation details, such as VAE decoding and KV-cache updates, actually account for more than half of end-to-end latency. The work replaces those components with lighter alternatives and speeds up diffusion sampling with systems optimizations.
Quality–latency comparisons are on the project page’s Pareto plot and examples. Apple says FastFrames-S preserves the visual quality of CF++ while being nearly 8× faster, and that FastFrames-L is comparable to Krea-Realtime on realism, prompt adherence, composition, and text rendering while being more than 10× faster. The 14B models, versus their 1.3B counterparts, are described as producing more realistic lighting and shadows, more coherent compositions, and more legible text. The page also says LingBot 1.3B optimized with FastFrames delivers a 16× speedup for fluid camera-conditioned control, and the demo compares RealWonder with physics-based dragging.
All of those figures are official reports under B200, 480p, steady-state conditions. This entry does not add license or hardware-support claims that were not taken from the project page. For deployment, use the license, dependencies, and hardware notes on the official page and Hugging Face model cards.
Usage
The first-party entry point is the project page FastFrames: Video Generation at Interactive Latency, with weights at apple/FastFrames-S and apple/FastFrames-L on Hugging Face. Officially they are autoregressive models for camera-conditioned or physics-conditioned video generation at interactive latency, not offline, high-latency long-form generators.
For size, S (1.3B) is the lighter model, officially 36 ms/frame on a single B200; L (14B) is the higher-quality model, with the official interactive figure of 45 ms/frame on 8×B200, or 158 ms on one GPU. Do not assume those B200 numbers will hold on consumer hardware or non-NVIDIA setups.
Cite them as FastFrames-S / FastFrames-L, or as the FastFrames family, so they are not confused with other FastFrame or video-acceleration projects. License, inference code, and dependencies should be taken from the model cards and project page as published.