DirCast: one streaming video model directs and commentates live Dota 2 broadcasts

Zixian Gao · hf · 2026-10-26

DirCast, from Alaya Lab, University of Tokyo, and Georgia Tech, is a single streaming video-language model that both directs live esports broadcasts (choosing cameras from raw feeds) and generates real-time commentary—using only past context.
Each second it scans all cameras at low resolution, decides via hold-or-switch whether and where to cut, then reads the chosen camera at high resolution before speaking; expiring views bound streaming cost.
Built on a dataset of 27 professional Dota 2 matches with per-hero cameras: it outperforms rule-based and zero-shot directors including GPT-5, and its commentary wins 29.2% of comparisons against GPT-4o versus 19.4% for the strongest of eleven streaming commentators.

More from Multimodal