arxivcs.CV2026-07-17
DiTango: Cost-Effective Parallel Diffusion Generation with Selective Attention State Reuse
Yuyang Chen, Runxin Zhong, Zan Zong, Hengjie Li, Yuyang Jin, Jidong Zhai
Recent advances in AI-generated content have driven widespread adoption of Diffusion Transformers (DiTs) for high-resolution, long-duration content generation. While parallelization techniques accelerate diffusion inference, they face significant scalability challenges due to exc…