SANA-Video 2.0 Hybrid Linear Attention for 720p Video Generation
July 22, 2026
SANA-Video 2.0 uses a hybrid architecture with gated linear attention and periodic gated-softmax anchors at a 3:1 ratio to achieve O(N) scaling. The 5B and 14B parameter models generate 720p video on a single GPU by using Block Attention Residuals to propagate feature updates.
HOW THIS AFFECTS YOU
●
builderYou can generate high-quality 720p video using significantly less compute via single-GPU architectures.
●
researcherYou can implement hybrid attention mechanisms to balance long-sequence scaling with full-rank token interactions.