SANA-Video 2.0 Generates 720p Video on a Single GPU in ~13 Seconds Using a 3:1 Linear-to-Softmax Attention Ratio

arXiv / HuggingFace Daily Papers·medium signal

SANA-Video 2.0 (arXiv 2607.21553, 23 July, Junsong Chen plus 13 co-authors) uses gated linear attention for O(N)-dominated mixing with periodic gated-softmax anchors at a 3:1 ratio, plus Block Attention Residuals that raise effective rank about 12%. The 5B and 14B models produce 720p on a single GPU: 13.06 seconds for a 720p 5-second clip fully optimized, 13.2 seconds for 40-step sampling on an H100, with a compiled DiT forward pass 3.2× faster than the full-softmax baseline at 720p and a VBench score of 84.30 at 480p. The authors claim roughly 120× the throughput of Wan 2.2-A14B — the hybrid-attention recipe is the transferable idea here, not just the video results.