Wan Streamer 0.3 gives interactive AI avatars full-body motion control

If you have spent any time experimenting with interactive AI avatars over the past year, you know their fatal limitation: they are essentially talking busts. They can match their lip flaps to audio, nod, and make occasional eye contact, but the moment you want them to do anything beyond sitting frozen in a webcam frame, the illusion falls apart.

Wan Streamer just introduced version 0.3, shifting the project from a portrait chat model into a dynamic character engine that moves, poses, and reacts to its environment.

While v0.1 and v0.2 perfected real-time gaze, micro-expressions, and conversational audio-video streaming, v0.3 introduces free-form physical behavior driven entirely by plain language. The research is detailed in a new technical paper titled "Video = World + Event Stream."

Splitting video into world and events

The fundamental challenge of real-time interactive video has always been temporal consistency: as soon as a character makes an abrupt move, their face melts or the room morphs into another place entirely.

Wan Streamer tackles this by decoupling video into two distinct layers: the world and the event stream.

The "world" encompasses everything that must stay rock-solid across time — character identity, room geometry, lighting, and overall visual aesthetic. The "event stream" handles everything that changes second by second, such as spoken dialogue, physical gestures, camera motion, and external happenings.

During a live session, the user's camera feed, microphone audio, and text input are fed straight into this event stream. The model then learns what happens next along the shared timeline, turning ordinary footage into rich pretraining data.

Actions in parentheses

In practice, v0.3 coordinates conversation and physical action by interleaving text cues right inside dialogue streams.

Behavior instructions appear in standard parentheses alongside spoken lines: "(reaches into the grass and picks up a green leaf) Look what I found" or "(covers mouth in surprise) I did not expect that." The brackets can contain virtually any natural-language description, from picking up objects and shifting posture to reacting to ambient sounds.

In published demos, users prompt the character with instructions like flashing a peace sign, putting hands on hips, fixing hair, glancing right, or looking down at the floor. The avatar executes the action fluidly while carrying on the verbal response.

Crucially, action and speech share a unified stream. The model learns and renders their timing simultaneously, bypassing the traditional need to stitch together separate speech pipelines and motion-capture rigs. The creators note that the live interaction demos reflect real end-to-end latency without sped-up footage or post-production compositing.

Why it matters

We think Wan Streamer 0.3 makes a brilliant architectural bet by treating physical action cues as primary stream tokens rather than downstream animation tasks.

Traditional digital humans rely on a cumbersome chain: an LLM outputs text, a voice model synthesizes audio, an animation module maps visemes or bones, and a renderer stitches the frames. That pipeline is brittle, latency-heavy, and visually artificial. Treating video generation as a unified renderer driven by an event stream points toward a genuinely interactive full-duplex future for virtual companions and game NPCs.

That said, you cannot download or deploy this on your local machine just yet. The team has only published the technical paper and demo results, with weights and real-time inference code still under wraps. Translating this architecture into cost-effective inference that runs outside high-end lab clusters will be the real hurdle to watch.