Video = World + Event Stream
Wan-Streamer v0.3 treats video generation as a live world state reacting to incoming events.
The paper defines the “world” as persistent context: scene, subjects, ambient sound, voices, and other stable conditions. The “event stream” is what changes over time, including behavior, speech, scene shifts, and sounds. The authors use that framing as a pretraining task for real-time video prediction and specialize it to full-duplex audio-visual interaction. They report the same operating point as v0.2: 640x368 video at 25 FPS, 160 ms streaming units, about 200 ms model-side response latency, and about 550 ms total interaction latency. HF Daily Papers' note
The paper defines the “world” as persistent context: scene, subjects, ambient sound, voices, and other stable conditions. The “event stream” is what changes over time, including behavior, speech, scene shifts, and sounds. The authors use that framing as a pretraining task for real-time video prediction and specialize it to full-duplex audio-visual interaction. They report the same operating point as v0.2: 640x368 video at 25 FPS, 160 ms streaming units, about 200 ms model-side response latency, and about 550 ms total interaction latency. HF Daily Papers' note
score 6