FastFrames-L: Apple's Two-Step Autoregressive Video Generation Model

Uploaded by the Apple Machine Learning Research team on 11 16. FastFrames is an autoregressive, two-step sampling video diffusion model that outputs 480p (832×480, 81 frames, 16 fps). The L version has 14B parameters: it first distills Wan2.1-14B into a two-step causal model using Causal Forcing++, then trains a pixel decoder and a KV correction network; the repository also includes all checkpoints for the causal teacher, consistency distillation, DMD distillation, and the final generator. There is also FastFrames-S with 1.3B parameters, as well as a LingBot 1.3B demo that is 16 times faster and supports camera control.

Original title: apple/FastFrames-L

Uploaded by the Apple Machine Learning Research team on 11 16. FastFrames is an autoregressive, two-step sampling video diffusion model that outputs 480p (832×480, 81 frames, 16 fps). The L version has 14B parameters: it first distills Wan2.1-14B into a two-step causal model using Causal Forcing++, then trains a pixel decoder and a KV correction network; the repository also includes all checkpoints for the causal teacher, consistency distillation, DMD distillation, and the final generator. There is also FastFrames-S with 1.3B parameters, as well as a LingBot 1.3B demo that is 16 times faster and supports camera control.

Why it matters For interactive video generation, the official team states that the smaller version takes only tens of milliseconds per step on a single GPU; the larger version is released along with checkpoints from all distillation stages

Rising score 42 · Likes 42 · Downloads 0

Reasons

  • Official latency breakdown for FastFrames-S on a single GPU at 480p: two DiT steps take about 16 ms each, KV correction takes 2.2 ms, and pixel decoding takes 0.9 ms, totaling 35.4 ms; speed tests for the L version use 8 GPUs
  • Weights for every distillation stage, from Wan2.1-14B to the final generator, are included in the repository, facilitating research reproduction
  • Includes VBench reproduction instructions and scripts for generating multiple videos on the project page 20

Usage

  • Linux x86-64, NVIDIA GPU, drivers supporting CUDA 12.8; clone GitHub apple-aiml-research/ml-fastframes, run uv sync, and download this repository's weights and third-party weights according to the checkpoint guide
  • Text-to-video: uv run python scripts/generate.py --model ff-l --checkpoint-dir checkpoints --prompt "..." --seed 0 --output out.mp4
  • Interactive demo: configure according to the lingbot documentation and run apps/lingbot/demo.py (use --gpus 8 for multi-GPU), then open localhost:8000 in your browser

Risks

  • The license is the Apple Machine Learning Research License (apple-amlr) and the Apple Sample Code License, with restricted usage; the LingBot portion is under CC BY-NC-SA 4.0, non-commercial only
  • The final DiT file is approximately 28.6 GB, and the entire repository including intermediate checkpoints approaches 200 GB; the official team has not specified single-GPU memory requirements, and both the L version demo and speed tests use 8 GPUs
  • The paper has not yet been released, there are no image quality comparison figures in the card, and output is limited to 480p; this entry has not been run

Open

Hugging Face Trending

Explore the latest AI Radar updates