Apple releases FastFrames video models for interactive 480p generation
Original title: apple/FastFrames-L — new model trending #30 on Hugging Face
FastFrames-S (1.3B) and FastFrames-L (14B) are two-step autoregressive diffusion models; an optimized LingBot 1.3B is 16x faster.
Apple published FastFrames-S (1.3B) and FastFrames-L (14B), autoregressive two-step video diffusion models that generate 480p video at interactive latency. Repository profiling lists 35.4 ms total latency for FastFrames-S on one GPU at 480p, versus 36 ms in the paper. FastFrames-L is distilled from Wan2.1-14B. The arXiv paper is marked coming soon. FastFrames code uses the Apple Sample Code License; LingBot-World v2 stays CC BY-NC-SA 4.0.
Unconfirmed: The arXiv paper is marked coming soon, so full published scores are not in this post.
- FastFrames-S has 1.3B parameters and FastFrames-L has 14B.
- Text-to-video output is 832 by 480, 81 frames at 16 fps.
- Profiled FastFrames-S latency is 35.4 ms on one GPU at 480p.
- Optimized LingBot 1.3B is described as 16 times faster than the official model.
- FastFrames-L training distills Wan2.1-14B into a two-step causal model.
- Vendors
- Apple
- AI models
- FastFrames-LFastFrames-SLingBot
From the original article
2 min read at huggingface.co[! [Project Page](https://img. shields.
By apple · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.