Token Radius Attention Speeds Up Video Generation in Diffusion Transformers Without Fine-Tuning

Researchers from Peking University have introduced Token Radius Attention (TRA), a method that reduces the amount of computation in Video Diffusion Transformers without requiring model retraining. The paper was published on arXiv on August 3, 2026. The authors observed that different tokens require different amounts of attention and proposed computing an individual interaction budget for each token based on entropy. The method was tested on seven configurations of Wan2.1, Wan2.2, and HunyuanVideo, where it retains 9 to 19 percent of attention interactions and delivers a 1.56–2.05x speedup with comparable quality.

What it means

The method directly concerns HunyuanVideo — one of the three tested models. Currently, HunyuanVideo requires at least 60 GB of VRAM for 720x1280 resolution and 129 frames, runs only on CUDA, and uses full 3D attention. TRA promises to reduce the computational load without changing the weights, which could potentially lower memory and generation time requirements. However, the tool's fact sheet does not include data on how much VRAM consumption would decrease with TRA — the method's developers do not specify these figures for specific configurations. The method's code is available in an open repository, allowing it to be integrated into existing HunyuanVideo pipelines. Since TRA does not require fine-tuning, owners of GPUs with less than the stated 60 GB of memory can test whether generation becomes feasible on their hardware. The AI video reference section will be updated once confirmed data on reduced memory requirements for specific models becomes available.