FastFrames: Video Generation at Interactive Latency
FastFrames’ 1.3B model generates a frame in 36 milliseconds on one B200 GPU, matching the quality of a model nearly eight times slower.
The researchers found that VAE decoding and KV cache updates accounted for more than half of end-to-end inference latency, then replaced them with lightweight alternatives. Their larger 14B model runs at 45 milliseconds per frame on eight B200 GPUs and delivers better visual quality and realism. The paper demonstrates real-time interaction in camera-conditioned and physics-conditioned video generation.
HF Daily Papers' note
The researchers found that VAE decoding and KV cache updates accounted for more than half of end-to-end inference latency, then replaced them with lightweight alternatives. Their larger 14B model runs at 45 milliseconds per frame on eight B200 GPUs and delivers better visual quality and realism. The paper demonstrates real-time interaction in camera-conditioned and physics-conditioned video generation.
HF Daily Papers' note
score 5