FastFrames generates video frames at interactive latency
Original title: FastFrames: Video Generation at Interactive Latency
A 1.3B variant runs at 36 ms per frame on one B200, matching a model nearly eight times slower.
The paper introduces FastFrames, autoregressive video diffusion models aimed at interactive latency. Profiling found VAE decoding and KV cache updates can exceed half of end-to-end latency, and the authors replace them with lighter alternatives. A 1.3B model reports 28 responses per second on one B200; a 14B model reports 22 responses per second on an 8x B200 node with higher visual quality.
- The 1.3B variant is reported at 36 ms per frame, or 28 RPS.
- The 14B variant is reported at 45 ms per frame, or 22 RPS.
- Demos include camera-conditioned and physics-conditioned video generation.
- Quality claims compare the 1.3B model with a nearly 8x slower baseline.
- Products
- B200
- AI models
- FastFrames
From the original article
1 min read at huggingface.coAutoregressive video diffusion models generate frames based on new observations and inputs, making them a promising medium for interactive visual experiences. However, existing models are orders of magnitude too slow for fluid interaction. This paper introduces FastFrames, a family of fast autoregressive video diffusion models that achieves interactive latency without compromising quality.
By Gene Chou, Guandao Yang, Fujun Luan · 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.