FastFrames: Video Generation at Interactive Latency

Project Page arXiv FastFrames-S FastFrames-L

This repository contains the official implementation of FastFrames: Video Generation at Interactive Latency. We release FastFrames-S (1.3B) and FastFrames-L (14B). Both are autoregressive, two-step video diffusion models that generate 480p video at interactive latency. We also release an optimized version of LingBot 1.3B that is 16x faster than the official model and allows fluid camera-conditioned control.

FastFrames demo: LingBot camera control with and without FastFrames optimizations

Installation

Linux x86-64 with an NVIDIA GPU and a CUDA 12.8-compatible driver. From the repository root:

uv sync

Download FastFrames and third-party weights using the checkpoint guide.

Latency Profiling

uv run python evaluation/latency/benchmark.py ff-s
uv run python evaluation/latency/benchmark.py ff-l --gpus 8

Example component timings for FastFrames-S, 1 GPU, 480p, with a full KV cache.

Component Calls Latency per call Total latency
DiT evaluation, t=1000 1 16.1 ms 16.1 ms
DiT evaluation, t=833 1 16.2 ms 16.2 ms
KV refinement 1 2.2 ms 2.2 ms
Pixel decoder 1 0.9 ms 0.9 ms
Total 35.4 ms (paper: 36 ms)

See the profiling guide for details and baselines.

Interactive Demo with LingBot

Complete the demo setup, then launch one of:

CUDA_VISIBLE_DEVICES=0 uv run --extra lingbot --extra demo python apps/lingbot/demo.py --gpus 1
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 uv run --extra lingbot --extra demo python apps/lingbot/demo.py --gpus 8

and open http://localhost:8000.

T2V Generation

uv run python scripts/generate.py --model ff-s --checkpoint-dir checkpoints \
    --prompt "A corgi surfing a wave at sunset" --seed 0 --output outputs/corgi.mp4

Outputs 832 Γ— 480 video, 81 frames at 16 fps. Use --model ff-l for FastFrames-L.

To generate the videos on our project page (20+ videos):

uv run python scripts/generate_project_page.py --model ff-s --output-dir outputs/project_page/ff-s
uv run python scripts/generate_project_page.py --model ff-l --output-dir outputs/project_page/ff-l

See the generation guide for details.

VBench Evaluation

Follow the evaluation guide to reproduce Table 1's Quality / Semantic / Total scores.

Training

Causal Forcing++ -> FastFrames-S

FastFrames-S keeps the two-step frame-wise Causal Forcing++ generator frozen and trains the networks around it.

Wan2.1-14B -> FastFrames-L

We first distill Wan2.1-14B into a two-step causal model following the Causal Forcing++ recipe, then train its pixel decoder and KV refinement network as for FastFrames-S. We release all checkpoints: 1) a causal teacher trained on latents generated by Wan2.1-14B; 2) the consistency-distilled model; 3) the DMD-distilled model; 4) the final generator, pixel decoder and KV refinement network.

Citation

If you find this work useful in your research, please cite:

@article{chou2026fastframes,
  title   = {FastFrames: Video Generation at Interactive Latency},
  author  = {Chou, Gene and Yang, Guandao and Luan, Fujun and Mescheder, Lars and
             Dong, Wei and Richter, Stephan R. and Koltun, Vladlen},
  journal = {arXiv preprint},
  year    = {2026}
}

License and Acknowledgments

LingBot-World v2 code and weights follow their upstream CC BY-NC-SA 4.0 license. All other FastFrames code is licensed under the Apple Sample Code License. Also see ACKNOWLEDGEMENTS for third-party code.

We thank the authors of Wan2.1, Self-Forcing, and Causal Forcing++ for releasing their code, on which the FastFrames code and models build.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for apple/FastFrames-L

Finetuned
(88)
this model

Space using apple/FastFrames-L 1