HeadCast makes 1080p video generation 1.95x faster

HeadCast cuts inference cost for autoregressive video diffusion by classifying every attention head once, at the point of maximum noise, before generation runs.

Attention math dominates inference cost in video diffusion, because the cache of past values keeps growing as a clip gets longer. Existing eviction heuristics fix that but cause flicker between frames. HeadCast avoids both, cutting inference 1.62x at 720p and 1.95x at 1080p, per the paper.

The classification happens once, sorting heads into four types: Sink, Dummy, Spatial, and Global. Each type then routes through its own cache pathway instead of one generic eviction rule.

Global heads get protected specifically, since they carry the long-range temporal consistency that keeps a video from stuttering. The Spatial pathway runs on a fixed-size grid, so savings grow as resolution climbs instead of staying flat.

VBench scores hold steady against full attention, so the speedup doesn't cost quality, at least by that benchmark.

Code is public at github.com/sjlgaga/HeadCast, so builders running autoregressive video diffusion can test the resolution-scaling claim themselves before committing inference budget to it.

The paper doesn't say how HeadCast performs above 1080p. It also doesn't say whether the four-way split transfers to video architectures outside the autoregressive diffusion family it was built for.