SANA-Video 2.0 hits 720p in 13 seconds on one GPU

SANA-Video 2.0 generates a 720p, five-second clip in 13.06 seconds on a single GPU, according to the paper posted to arXiv. That's the fully optimized number; the compiled DiT forward pass alone runs 3.2 times faster than a full-softmax build at the same resolution.

That speed matters because attention is the expensive part of video diffusion transformers, and the cost climbs fast with sequence length, exactly what 720p frames produce. SANA-Video 2.0 handles most of that mixing with gated linear attention, an O(N) approach, and drops in gated-softmax attention as an anchor every fourth block, a 3:1 ratio. The authors add what they call Block Attention Residuals on top, which raise the model's effective rank by about 12%.

Both the 5B and 14B versions fit on a single GPU. At 480p the model scores 84.30 on VBench, and the paper claims roughly 120 times the throughput of Wan 2.2-A14B, a full-softmax model in the same weight class.

The paper's own framing treats the video results as secondary to the method. The 3:1 linear-to-softmax ratio and the rank fix are architecture choices any diffusion transformer team can lift, video generation or not. Watch for other labs to start reporting their own linear-to-softmax ratios instead of just parameter counts.