Raw capture — raw/sana-video-2.md

url: https://nvlabs.github.io/Sana/Video2/ title: "SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation" authors: Junsong Chen, Jincheng Yu, Yitong Li, Shuchen Xue, Haozhe Liu, Jingyu Xin, Yuyang Zhao, Tian Ye, Zhangjie Wu, Zian Wang, Daquan Zhou, Ping Luo, Song Han, Enze Xie date_fetched: 2026-07-25 date_published: 2026 affiliation: NVIDIA Research (Efficient AI Team & Singapore Lab) arxiv: 2607.21553 doi: 10.48550/arXiv.2607.21553

SANA-Video 2.0 — Hybrid Linear Attention with Attention Residuals for Efficient Video Generation

Authors

Junsong Chen, Jincheng Yu, Yitong Li, Shuchen Xue, Haozhe Liu, Jingyu Xin, Yuyang Zhao, Tian Ye, Zhangjie Wu, Zian Wang, Daquan Zhou, Ping Luo, Song Han, Enze Xie

Affiliation: NVIDIA Research (Efficient AI Team & Singapore Lab)

Publication Details

• eprint: 2607.21553 (arXiv, primaryClass cs.CV)
• Year: 2026
• DOI: 10.48550/arXiv.2607.21553

Links

• Paper on arXiv
• Code on GitHub

Key Metrics (callout banner)

• 84.30 VBench Total
• 13.06s for 720p/5s on one H100
• 3.2× faster than softmax at 60s
• 120× faster than Wan 2.2 14B

One-H100 Latency Benchmarks (720p / 5s / 40 steps)

(Note: log scale, lower is better)

Abstract & Technical Details

The paper introduces a hybrid video diffusion transformer at two scales (5B and 14B) under a unified architecture, targeting high-quality 720p video generation on a single GPU. The system matches full-softmax video DiTs in quality while benefiting from linear attention's favorable long-sequence scaling.

Hybrid Linear-Softmax Attention

The method "combines gated linear attention for O(N)-dominated mixing with periodic gated-softmax anchors at a 3:1 ratio." This restores full-rank token interactions that pure linear attention lacks, avoiding quadratic attention throughout.

Block Attention Residuals (AttnRes)

Completed block summaries are routed into later linear layers, enabling anchor-feature reuse. This boosts deep-layer effective rank by approximately 12%.

Training Approach

The model is trained from scratch — the complete hybrid is learned directly rather than linearizing pretrained models. Reduced-resolution proxy studies identified 25% softmax as the optimal quality-efficiency trade-off.

Performance Highlights

• At 40-step sampling, SANA-Video 2.0 achieves a VBench score of 84.30 in 13.2 seconds at 480p on a single H100, remaining competitive with far larger softmax video DiTs.
• The compiled DiT forward pass is "3.2× faster than a matched full-softmax baseline at 720p/60s," with the gap widening as video duration increases.
• Full-stack Sol-Engine optimization (kernel fusion, caching, sparse attention) accelerates the backbone by an additional 3.58×, bringing the 5B pipeline to 13.06s at 720p/5s.
• Overall, the system is 120× faster than Wan 2.2-A14B on one H100.

Bottom-Line Claim

The authors state their "hybrid design recovers softmax-level expressiveness at substantially reduced cost, unlocking scalable long, high resolution video generation."

Additional Notes

• The page notes "VIDEO SOURCE COMING SOON" for a video showreel section.
• BibTeX citation is provided on the page.