SPADE accelerates attention in video diffusion transformers without fine-tuning

Researchers have presented SPADE, a sparse attention engine that accelerates inference for video diffusion transformers without model fine-tuning. The paper was published on August 4, 2026 on arXiv. The authors propose three components: the vDiT-SSR specification for 3D token blocking and dynamic masks, a runtime scheme with per-frame policy, and an executor with block-sparse flash attention. On Hunyuan-Video and Wan 2.1/2.2 models, attention speedup ranged from 2.26x to 3.40x, and end-to-end inference from 1.49x to 1.80x. The code is open source on GitHub.

What it means

The news directly concerns the Wan generator: the authors tested SPADE on Wan 2.1 and 2.2 for text-to-video and image-to-video tasks. The Wan fact sheet states that the T2V-1.3B model generates a 5-second 480p video on an RTX 4090 in about 4 minutes without optimizations like quantization. SPADE is exactly such an optimization, but at the attention mechanism level, and it does not require changing the model weights. Wan's developer, Alibaba, is not involved in this work — SPADE was created by an independent group. A 1.49x—1.80x speedup in end-to-end inference means the same video can be generated faster with the same hardware requirements. How much faster exactly on consumer GPUs like the RTX 4090, the authors do not specify — the numbers were obtained in their test environment. SPADE is a method, not a standalone product. To use it, you need to integrate the code from the repository into the Wan inference pipeline. The developer does not provide a ready-made build or GUI. In our video generator section, Wan remains with the same specs as before, but now technically savvy users have a way to cut generation time without switching models.