NVIDIA Research has published SANA-Video 2.0, a video generation architecture that reaches 720p output on a single H100 GPU in roughly 13 seconds, cutting the hardware footprint that high-resolution AI video usually demands. The paper is credited to a team led by Junsong Chen inside NVIDIA’s Efficient AI Team and its Singapore lab. It targets the practical barrier most video diffusion systems run into: multi-GPU clusters required to render more than a few seconds of footage.

The core mechanism is what NVIDIA calls Hybrid Linear-Softmax Attention. Most layers use linear attention, which scales cheaply with sequence length, while roughly one in four layers reverts to standard softmax attention to restore the fine-grained token interactions that linear attention alone loses. NVIDIA’s own proxy studies, run at reduced resolution, found a 3

ratio of linear to softmax layers to be the best trade-off between speed and quality.

A second mechanism, Block Attention Residuals, carries completed block outputs forward into later linear-attention layers so the network can reuse full-rank information deeper in the stack. NVIDIA reports this lifts effective rank in deeper layers by about 12 percent, a proxy for how much distinct visual information the model retains as a clip gets longer.

The model was trained from scratch as a unified hybrid rather than adapted from an existing softmax-only video model, a choice NVIDIA says was needed to learn the combined attention pattern properly. It ships in two sizes, 5 billion and 14 billion parameters, sharing one architecture.

On VBench, a standard video-quality benchmark, the 5-billion model scores 84.30 in 13.2 seconds at 480p using 40 sampling steps on one H100, which NVIDIA says holds up against much larger softmax-based video models running at far higher latency. Layering in NVIDIA’s Sol-Engine, a separate optimization stack that fuses kernels, caches computation, and applies sparse attention, brings 720p generation down to 13.06 seconds on one H100. NVIDIA’s own chart puts Wan 2.2’s 14-billion-parameter model at 1,556 seconds on the same hardware, a gap it labels 120 times.

That comparison is NVIDIA’s own, run on its own hardware and its own benchmark protocol. Independent replication has not been published. Still, the constraint it addresses is real: renting a multi-GPU node for every video request is a cost few products can absorb at scale. A model that turns out a 720p clip on one GPU in seconds rather than minutes changes the math for latency-sensitive products, including social video tools, ad-creative generators, and game asset pipelines serving many concurrent users without a dedicated cluster per request.

Teams evaluating video infrastructure for the second half of 2026 should treat the 120-times figure as a ceiling set under NVIDIA’s own test conditions. Benchmark SANA-Video 2.0 against Wan 2.2 and LTX-2.3 on the GPUs already in your own stack before assuming the reported gap survives outside a single H100.

NVIDIA Research published the SANA-Video 2.0 project page and its accompanying arXiv paper, posted in July 2026, credited to lead author Junsong Chen and coauthors from NVIDIA’s Efficient AI Team and Singapore lab.