A research team has pushed Wan-Streamer to higher video quality without making users wait longer for a response.
Wan-Streamer v0.2 bumps the interactive output resolution from 192x336 to 640x368 — more than triple the pixel count — while keeping model-side latency at roughly 200 ms at 25 FPS. Total round-trip latency, including a 350 ms network budget, sits at about 550 ms. The team achieved this by splitting work across two components: a single-GPU "thinker" handles streaming perception and builds a generation cache, while a multi-GPU "performer" group — arranged in what the paper calls Ulysses-style context parallelism — handles the heavy lifting of generating high-resolution video latents. Audio generation skips the sequence-sharding step entirely, keeping that path lean.
The resolution jump matters because it makes on-screen agents legible in ways the earlier version could not support — posture, gaze, hands, and surrounding objects become readable during live conversation, which is the difference between a blurry avatar and something closer to a usable interface. Real-time audio-visual AI interaction has been a crowded research target, and the central engineering challenge has always been the same: resolution and latency pull in opposite directions, and every team claiming to have squared that circle deserves scrutiny.
This one at least shows its math — the architectural split between a lean thinker and a parallelized performer is a concrete mechanism, not a marketing claim. Whether 550 ms total latency feels responsive enough for real interaction is a question the paper does not answer.