The authors build on diffusion transformers that generate video in successive chunks.
Learned attention proxies help streaming video generators retain historical context
Gao and colleagues report improved short- and long-horizon video generation by training compact representations to approximate attention responses from earlier frames.
Chinese Tech
Junyao Gao · Sibo Liu · Jiaxing Li · Yanan Sun · Hanxin Zhu · Weidong Zhang · +2 more
Tongji University · Tencent · Nanyang Technological University · Hong Kong University of Science and Technology
Research Digest··2 min read
Streaming video generators often discard older context to keep memory and computation manageable, sacrificing information that could guide later frames.
§