Skip to content
AI.info

The Pulse

Apple releases FastFrames-L video model for research use

FastFrames-L is a 14-billion-parameter text-to-video model released alongside a smaller FastFrames-S model. Apple says both generate 480p video at interactive latency, and the model license limits use to research.

Apple releases FastFrames-L video model for research use

Apple has released FastFrames-L, a 14-billion-parameter text-to-video model, alongside a smaller 1.3-billion-parameter version called FastFrames-S. Both are autoregressive, two-step diffusion models that generate 480p video at interactive latency, according to the company’s Hugging Face model card.

The release includes model checkpoints, generation and evaluation scripts, and an interactive demonstration. The model license limits use to research and excludes commercial products and services.

How the models were built

For FastFrames-L, Apple distilled Wan2.1-14B into a two-step causal model using the Causal Forcing++ training recipe, then trained a pixel decoder and a key-value refinement network. The released checkpoints include a causal teacher trained on latent representations generated by Wan2.1-14B, a consistency-distilled model, a DMD-distilled model and the final generator with its accompanying networks.

For FastFrames-S, the company kept the two-step, frame-wise Causal Forcing++ generator frozen and trained the networks around it. The model card identifies Wan2.1, Self-Forcing and Causal Forcing++ as earlier projects on which the FastFrames code and models build.

Generation and timing

An example video-generation command for FastFrames-S produces video at 832 by 480 pixels, with 81 frames at 16 frames per second. The same generation script can be configured to use FastFrames-L. A separate script is provided to generate the more than 20 videos on the project page.

The setup requires Linux x86-64, an NVIDIA GPU and a CUDA 12.8-compatible driver, along with FastFrames and third-party weights. The release also provides commands for latency profiling and instructions for reproducing quality, semantic and total scores using VBench.

Apple provides an example timing breakdown for FastFrames-S at 480p on one GPU with a full key-value cache: the components total 35.4 milliseconds. The two diffusion transformer evaluations account for 16.1 and 16.2 milliseconds, while key-value refinement takes 2.2 milliseconds and the pixel decoder takes 0.9 milliseconds. The model card gives those figures for the smaller model, not FastFrames-L.

Demo and license terms

The release also includes an optimized LingBot 1.3B model for camera-conditioned control. Apple says that version is 16 times faster than the official model and allows fluid camera-conditioned control; the interactive demo can be started with one or eight GPUs.

LingBot-World v2 code and weights retain their upstream CC BY-NC-SA 4.0 license. Other FastFrames code is licensed under the Apple Sample Code License.

Sources

  1. Apple
  2. Apple

Related stories

Explore

More articles