Meta announced Muse Realtime Avatar on September 23, 2026. It turns Muse Realtime Voice into an expressive, interactive avatar you can talk to live. Here's the whole pipeline on one page.
The pipeline in 4 steps
- Voice makes speech tokens. Muse Realtime Voice produces a stream of speech tokens (VQs) that carry what is said and how it's delivered.
- The avatar reads the same tokens. Voice, lip motion and expression all come from one shared token stream, so they stay in sync.
- Video is generated in short chunks. A Diffusion Transformer is conditioned on the speech tokens, the reference media, and a rolling window of recent video latents. It generates video in short causal chunks.
- A distilled "student" makes it fast. The student reproduces in 2 unguided evaluations what the teacher refines over 120 guided ones. That's a 60x cut in compute per chunk.
The specs
| What | Number |
|---|---|
| Latency (end of your turn → first byte of voice + video) | ~870 ms |
| Video | 448×768 portrait, 25 fps |
| Model time per frame | 2.5 ms |
| Live sessions per GB200 GPU | 12 |
What it can animate
Meta says it goes "beyond talking heads": any image, with facial, hand and full-body movement. A photo portrait responds with subtle expressions, a full-body illustration gestures and shifts posture, and animals and everyday objects become expressive too.
How it compared (Meta's own study)
Raters held 2–3 minute conversations with each system, using matched avatar identities.
| Muse Realtime Avatar vs | Preferred Muse |
|---|---|
| Runway Characters | 78% vs 22% |
| HeyGen LiveAvatar | 88% vs 12% |
Things to keep in mind: Meta ran this study itself, and the mannerisms score against Runway was statistically a tie.
Good to know
- Watermarked: generated video carries an invisible Meta Video Seal watermark.
- Adults only: Muse is for users 18 and older.
- Demos vs the app: Meta says not every avatar in its examples is available in the Muse app yet. TechCrunch says video chat with your own avatar is coming "soon".
Why it matters
The voice and the video come from one token stream. That's how it can answer in under a second with voice, lips and expression in sync.
Sources: research.meta.ai/blog/bringing-your-muse-to-life · techcrunch.com · runtimewire.com