Xingyang Li, Muyang Li, Tianle Cai, Haocheng Xi, Shuo Yang, Yujun Lin, Lvmin Zhang, Songlin Yang, Jinbo Hu, Kelly Peng, Maneesh Agrawala, Ion Stoica, Kurt Keutzer, Song Han
Radial Attention is a sparse attention mechanism for video diffusion models that improves efficiency and speed by leveraging spatiotemporal energy decay, achieving significant performance gains in video generation tasks.
Generating long videos using advanced AI models can be very resource-intensive due to the extra time dimension involved. This paper introduces a new method called Radial Attention, which mimics how signals naturally lose strength over distance, to make the process more efficient. By focusing computational power on nearby video frames and reducing it over time, this method allows for faster video generation without sacrificing quality. This approach also enables existing AI models to generate longer videos with much less training and processing time.
