Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)

Apple Makes a Video Model 7.6x Faster: 36 ms per Frame Instead of 272 ms

Video models that generate a clip frame by frame respond to a keypress with a noticeable pause in games and simulators. The best open model, Causal Forcing++, spent 272 ms per frame on a B200 GPU. Apple's FastFrames method cuts that to 36 ms (the S version), which is 7.6x faster. The trick is replacing the heavy VAE decoder (147 ms per frame) with a mini-network that takes 1 ms. People start to notice latency at 50–60 ms, so the goal is to stay within 50 ms per frame. FastFrames-L, with 14 billion parameters, reaches 45 ms per frame on eight B200s. On VBench, FastFrames-S scores 84.23 against 84.11 for Causal Forcing++. The weights are released under the Apple Model License for Research, which allows research use only.

AI-processed from TG: @ai_machinelearning_big_data (Machinelearning); edited by Hamidun News
Apple Makes a Video Model 7.6x Faster: 36 ms per Frame Instead of 272 ms
Source: TG: @ai_machinelearning_big_data (Machinelearning). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @ai_machinelearning_big_data (Machinelearning) ↗
Editor-in-Chief
FastFrames-L with 14 billion parameters45 ms per frame on eight B200s
VBench84.23 points for FastFrames-S vs. 84.11 for Causal Forcing++
LicenseApple Model License for Research, research use only

Imagine a game in which a neural network draws every frame: you press "forward," and the world reacts a full second later. That kind of delay keeps video models that generate clips frame by frame from becoming the basis for games, simulators and predictive models for robots. Apple has released FastFrames, a method that shortens this pause.

The fastest open model of this type, Causal Forcing++ with 1.3 billion parameters, spent 272 ms per frame on an Nvidia B200 GPU. The FastFrames-S version produces the same frame in 36 ms, which is 7.

6x faster. The benchmark to keep in mind: people notice latency from roughly 50–60 ms, so each frame should take no more than 50 ms. The practical takeaway for anyone building interactive video: measure the latency of the response to an action, not frames per second.

Why FPS is misleading here

The usual speed metric is frames per second (FPS). But it shows average throughput, not responsiveness. Suppose a model generates video in batches of 30 frames and spends one second on each. On paper, that is 30 FPS. But a player who presses a key right after a batch starts will see the result only a second later, when the next block is ready.

So Apple measures the delay from a command to the first frame that shows its effect. The target threshold comes from human-computer interaction research: people already notice delays above 50–60 ms. That means each frame should take no more than 50 ms.

Where the time went and what was replaced

To find where the milliseconds were being lost, the team profiled Causal Forcing++. On a single B200 it spends 272 ms per frame, and only about 30% of that time goes to diffusion steps, the process of turning noise into an image.

The main losses were elsewhere:

  • The VAE decoder, which converts the frame's internal (latent) representation into pixels, takes 147 ms, more than half the time.
  • Another 41.6 ms goes to an extra transformer pass whose only job is to update the KV-cache, the model's memory of what it has already shown.

FastFrames replaces both components. In place of the VAE, a lightweight decoder made of three convolutional blocks and a linear layer takes 1 ms per frame. In place of the extra transformer pass, a small network spends 2 ms correcting the cache that was written from the noisy input. The transformer itself is not changed or retrained. The rest of the speedup comes from system-level optimizations, including distributing computation across several GPUs.

Numbers and models

Apple has published two models on Hugging Face, both based on Wan 2.1: FastFrames-S with 1.3 billion parameters and FastFrames-L with 14 billion. Both generate video from a text description: clips of about five seconds at 832×480 resolution, 81 frames, 16 frames per second.

FastFrames-S produces a frame in 36 ms on a single B200, 7.6x faster than Causal Forcing++. Quality did not drop as a result: on the VBench benchmark the model scores 84.23 against 84.11.

FastFrames-L fits within 45 ms per frame on eight B200s and takes 158 ms on a single GPU. Its VBench score is 84.44, on par with Krea-Realtime 14B (84.42), which needs 1675 ms per frame.

The method also carries over to other teams' models. LingBot, an open world model from Robbyant that you can walk around in with a keyboard, now responds to a keypress in 52 ms instead of 842. RealWonder, a model with physical control, responds in 51 ms instead of 1141. Both, however, need eight GPUs to do it.

Limitations and what to do in practice

There are several caveats. The record figures for the large models require eight B200-class GPUs, so this is not something for a laptop or an ordinary cloud instance. The 50 ms threshold is reached only on the most expensive configurations, and FastFrames-L on a single card, at 158 ms, falls short of it. In addition, the weights are released under the Apple Model License for Research, which covers research use, not commercial products.

For developers of interactive video systems, such as game prototypes, world simulators and predictive models for robots, the takeaway is simple: measure not FPS but the time from keypress to visible reaction. And look for bottlenecks beyond the diffusion steps: here the main culprits were the decoder and cache maintenance. Replacing them with small trained networks delivered almost all of the gain without retraining the main model.

The signal for product teams is also clear: video models are moving from a "generate and show" mode to a "respond to input" mode. For now this is a research release, but the open weights let you verify the results yourself.

Story

Key Questions Answered

Why FPS is misleading here?

The usual speed metric is frames per second (FPS). But it shows average throughput, not responsiveness. Suppose a model generates video in batches of 30 frames and spends one second on each. On paper, that is 30 FPS. But a player who presses a key right after a batch starts will see the result only a second later, when the next block is ready. So Apple measures the delay from a command to the first frame that shows its effect.

Where the time went and what was replaced?

To find where the milliseconds were being lost, the team profiled Causal Forcing++. On a single B200 it spends 272 ms per frame, and only about 30% of that time goes to diffusion steps, the process of turning noise into an image. The main losses were elsewhere: The VAE decoder, which converts the frame's internal (latent) representation into pixels, takes 147 ms, more than half the time.

What are the key details on Numbers and models?

Apple has published two models on Hugging Face, both based on Wan 2.1: FastFrames-S with 1.3 billion parameters and FastFrames-L with 14 billion. Both generate video from a text description: clips of about five seconds at 832×480 resolution, 81 frames, 16 frames per second. FastFrames-S produces a frame in 36 ms on a single B200, 7.6x faster than Causal Forcing++. Quality did not drop as a result: on the VBench benchmark the model scores 84.23 against 84.11.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…