NVIDIA has released code that makes existing video generators roughly twice as fast without retraining them. On July 28 the company landed Sol-Attn in the sol-engine branch of its Sana repository, integrating it into HunyuanVideo-13B and Wan2.1-T2V-14B. The technique reports up to 2.1 times faster video generation and 2.3 times faster video editing while preserving visual quality, and it changes no model weights at all.
Key facts
- The headline number: up to 2.1 times faster video generation and 2.3 times faster editing from the technique alone, rising to up to 5 times only when bundled with NVIDIA's other Sol-Engine optimizations.
- When: code released July 28, 2026; the paper was submitted July 27.
- Who: NVIDIA researchers Haopeng Li, Yitong Li, Junsong Chen, Tian Ye, Haozhe Liu, Jincheng Yu, Duomin Wang, Ruihua Zhang, Zeke Xie, Enze Xie and Song Han.
- Primary source: the sol-engine branch on GitHub, with the project page and paper.
Video generation is expensive for a specific, structural reason. The models are diffusion transformers, and attention -- the mechanism that lets every piece of the video consider every other piece -- grows quadratically with sequence length. A few seconds of high-resolution video is an enormous number of tokens, so attention dominates the cost. Our explainer on sparse attention covers the standard escape route: compute only the parts that matter and skip the rest.
The trouble is that deciding what to skip is itself expensive. Existing training-free methods first build a routing map -- scoring blocks of the sequence with a cheap proxy, then selecting either a fixed fraction of the top-ranked blocks or enough of them to reach a probability threshold. Computing and storing those proxy scores costs real time and memory, and fixed budgets are either wasteful or too aggressive depending on the content.
Sol-Attn's move is to fold the decision into the computation. As attention runs its normal streaming pass over the sequence, it screens compressed key and value blocks on the fly: promising blocks get exact attention, and the skipped ones contribute an approximate correction rather than being dropped entirely. Routing, sparse computation and correction all happen inside one kernel, so there is no separate routing map to materialize.
A useful analogy is proofreading a long manuscript. The naive approach reads every page with equal care. The typical sparse approach first skims the whole thing to rank which pages need attention, which costs a full extra pass. Sol-Attn decides page by page as it reads, and leaves a short note where it skimmed so nothing is silently lost.
What makes this immediately practical is what it does not require. There is no distillation, no fine-tuning, no step-count reduction, no change to the sampling procedure, and no new weights -- so unlike diffusion distillation, which trades quality characteristics for speed, this is a substitution underneath an unchanged model. The project reports results across several generators and one video editor, and publishes side-by-side comparisons using identical prompts and seeds so the quality claim is at least inspectable.
Be precise about the numbers, though, because two very different figures are circulating. The 2.1 and 2.3 times results are what Sol-Attn delivers in NVIDIA's tested pipelines. The "up to 5 times end-to-end" figure requires the rest of the Sol-Engine stack -- kernel fusion, caching, low-precision formats, token pruning -- and attributing that to Sol-Attn alone would be wrong. Notably, the repository's own model table marks the two pipelines Sol-Attn just landed in as "re-benchmark pending," which is an unusually candid admission that the integrated end-to-end numbers are not final.
The honest caveats: every performance figure here is author-reported on NVIDIA hardware with warmup excluded, no independent reproduction exists yet, and approximate attention always carries some risk of quality regression that side-by-side demos may not expose -- particularly on content types the authors did not test. For anyone running local video generation, the payoff is nonetheless concrete and free: the same model, the same weights, roughly half the wall-clock time. That continues a pattern Ground Truth has tracked all month, in which the local inference floor keeps falling through kernel work rather than new models.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)