SANA-Video 2.0 Runs 720p on One GPU With 3.2x DiT Speedup

A new paper from the SANA-Video team quietly moves the goalposts on what one GPU is supposed to be able to do in video generation, and the interesting bit is how it gets there, not the leaderboard number at the top of the page.

The arXiv preprint describes SANA-Video 2.0 at 5B and 14B parameters, generating up to 720p video on a single GPU. The headline efficiency claim is a 3.2x speedup of the DiT forward pass at 720p versus a softmax baseline, with an additional 3.58x acceleration from what the authors call Sol-Engine, ending at roughly 13.06 seconds for a 720p 5-second clip on H100 with the 5B model. They also report a VBench score of 84.30 and claim their system runs about 120x faster than Wan 2.2-A14B on the same hardware.

The recipe underneath is worth pausing on. Instead of pure softmax attention or pure linear attention, the model uses a hybrid, with gated linear attention interleaved with periodic gated-softmax anchors at a 3:1 ratio. Linear attention gives O(N) scaling that makes long video sequences tractable, and the softmax anchors reportedly preserve the full-rank token interactions that pure linear variants tend to lose. Layered on top is what they call Block Attention Residuals, routing completed summaries into deeper layers and improving effective rank by about 12%. The paper's claim is softmax-level expressiveness at substantially reduced cost.

Why this matters if you are not training video models yourself: the class of workloads that used to need multi-GPU inference to move at all, 720p diffusion transformers for video, starts to fit on one card, which changes the cost picture for anyone fine-tuning or serving one of these systems. Small studios, research groups without a cluster, and infrastructure teams sizing GPUs for video features all get a different bill.

Take the specifics as reported, not settled. VBench is one benchmark and a 5-second 720p H100 clip is a favorable stress test, and what the paper does not give you is behavior on longer, high-motion or text-heavy prompts, whether the 14B weights will actually ship publicly, or how the efficiency numbers translate off H100-class hardware. The direction, though, hybrid linear attention displacing pure-softmax DiTs on the efficiency curve, is the part worth watching, because it is the same trick that already reshaped LLM inference now arriving in video.