Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-23 · high
It classifies every attention head once per clip into four roles, skips fine-tuning, and matches full-attention quality on VBench.
Why now: The paper's arXiv identifier, 2607.20125, places it in July 2026, with runnable code already posted alongside it for testing the resolution-scaling claim.
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.
Each link below shares sources, entities, or timing with this story.
Shared entity: Code / Same source domain / Earlier coverage / Tension
Both cover Code; reported by the same outlet (arxiv.org); earlier Code coverage from 2026-06-15.
Shared entity: Code / Shared topic / Earlier coverage / Tension
Both cover Code; overlapping topics (code, cost); earlier Code coverage from 2026-06-01.
Same source domain / Shared topic
Reported by the same outlet (arxiv.org); overlapping topics (attention, autoregressive, code, diffusion, video).
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (autoregressive, code, diffusion); pushes against this story (but).
Same source domain / Shared topic
Reported by the same outlet (arxiv.org); overlapping topics (attention, cost, diffusion, video).
Reported by the same outlet (arxiv.org); overlapping topics (attention, cost, diffusion, video).
Shared entity: Code / Same source domain / Earlier coverage
Both cover Code; reported by the same outlet (arxiv.org); earlier Code coverage from 2026-05-29.
Shared entity: Code / Shared topic / Earlier coverage
Both cover Code; overlapping topics (code, cost); earlier Code coverage from 2026-05-19.