arxivcs.CVcs.LG2026-07-22
HeadCast: Casting Attention Heads for Efficient Autoregressive Video Generation
Jinliang Shen, Lianghao Su, Zheming Li, Kang He, ZiLiang Lai, Yanbing Jiang, et al.
Autoregressive (AR) video diffusion models have become a promising paradigm for long and streaming video synthesis, but the continuously growing Key-Value (KV) cache makes attention the dominant inference cost, especially at high resolution where each frame contributes many token…