A new GPU runtime called Tessera ran a 50-step video-generation loop 1.22 to 2.08 times as fast as two comparison systems in tests on an NVIDIA H100. The researchers measured that range across two 720p video models, using FlashInfer and flex-block-attn as the baselines. The improvement came from changing how the GPU executes sparse attention, a costly part of each generation step.
Video diffusion models build a clip by repeatedly refining a representation of its frames. Attention lets one part of that representation draw on another, but computing every possible interaction is expensive. Block-sparse attention skips groups of interactions selected by a mask. As the mask changes between requests, a fixed way of dividing the work across the GPU can waste time moving data or leave computing units underused.
Tessera keeps the mask's chosen interactions intact while separating those logical groups from the physical tiles the GPU processes. It can keep a group at its original size, combine groups into larger tiles, or split them into smaller ones. Larger tiles can reuse data but may calculate over masked-out entries; smaller ones can avoid that work but move more data. The runtime also groups and schedules tiles, then uses rankings built by offline profiling to select a precompiled plan for each request without searching through plans on the critical path.
In a separate benchmark of 2,315 video-attention masks on A100, H100, RTX 4090 and RTX 5090 GPUs, the authors report lower latency for complete sparse-attention requests than the tested alternatives in all 76 measured platform, baseline and workload-family combinations. For the two H100 video-model runs, sparse-attention requests were 2.95 to 6.79 times as fast as the comparison systems. The smaller 1.22-to-2.08-times gain for the full 50-step loop reflects computation outside attention that Tessera does not change.
The model-level tests used Wan2.1 text-to-video and HunyuanVideo-1.5 image-to-video at 720p on an H100 SXM. Those measurements cover the diffusion loop, rather than the latency of a complete video-generation service. Tessera is described in a new preprint; its speedups are the researchers' own results on the tested sparse masks, models and NVIDIA GPUs.
Sources
- Research paper and submission record, arXiv:2609.25869
- Full paper, including evaluation methods and results
- Illustrative image: NVIDIA H100 video still, Geekerwan, via Wikimedia Commons, CC BY 3.0. The pictured PCIe card is not the H100 SXM used in the study.
_025.png){kind=link}