1 paper
Cheng Liang, Haoxian Chen, Liang Hou +4
The quadratic complexity of the attention mechanism severely limits the context scalability of Video Diffusion Transformers (DiTs). We find that the highly sparse spatio-temporal a…