SANA-Video 2.0: Hybrid Linear Attention Enables 720p Video Generation on a Single GPU

SANA-Video 2.0：混合线性注意力与注意力残差实现高效视频生成

This paper replaces standard attention in video diffusion with a 3:1 hybrid: gated linear attention for most tokens, periodic softmax anchors to restore full-rank interactions. The 5B model hits VBench 84.30 at 480p in 13.2s on one H100. With full-stack Sol-Engine optimizations, 720p/5s generation takes 13.06s—120x faster than Wan 2.2-A14B on the same GPU. Block Attention Residuals route early-layer summaries to deeper layers, boosting effective rank by ~12%. The post does not disclose open-source plans or training data details.

Why it matters: Concrete, reproducible architectural change hits H and K, but the pure-paper format lacks a product hook so R is absent. Lands right at the featured threshold of 72 per policy, no bump applied.