Nvidia's New Long-Form Video Generation Technology

Nvidia introduced SANA-Video 2.0, a hybrid video diffusion transformer capable of generating high-quality 720p video on a single GPU. The model is implemented at 5B and 14B scales, securing both the efficiency of Linear Attention and the quality of Full-softmax attention at the same time.

The core technology, Hybrid Linear-Softmax Attention, combines gated linear attention with periodic gated softmax anchors at a 3:1 ratio, resolving the lack of token-to-token interaction that is a limitation of linear attention while lowering computational complexity. In addition, Block Attention Residuals (AttnRes) reuse anchor features to raise the Effective Rank of deep layers by about 12%.

The key performance characteristics are as follows:

• Overwhelming speed: At 720p/60s, it achieves a DiT forward pass that is 3.2x faster than the existing Full-softmax baseline.
• Hardware optimization: Through Sol-Engine optimization, it builds a hardware-friendly backbone, generating a 720p/5s video in 13.06 seconds on the 5B pipeline, recording a speed 120x faster than Wan 2.2-A14B.
• High quality: It achieves a VBench score of 84.30 at 480p, demonstrating efficiency that can compete with much larger-scale models.