Meta AI Research unveiled Muse Realtime Avatar on September 23, a system that takes a single reference image, a portrait, a cartoon, even a photo of an animal or an object, and animates it into a live, talking video character while a person speaks to it. The avatar moves its face, hands, and body in sync with generated speech, frame by frame, for as long as the conversation runs.
The release matters because it turns Meta’s existing voice assistant, Muse Realtime Voice, into something with a face. Meta pairs the two into one streaming pipeline: the voice system encodes what is said and how it sounds into a single token stream, and the avatar model reads that same stream to build matching video. Sharing one token stream is what keeps the mouth, hands, and expression lined up with the audio rather than drifting apart over a long exchange.
Under the hood, Meta trained a “student” model to imitate a much slower “teacher” model. The teacher needs 120 model evaluations to render each short video chunk. Meta’s distilled student needs two, a 60 times reduction, while Meta says it preserves close to the same visual quality raters saw from the teacher. That kind of compression is what makes a system like this affordable to run at the scale of a consumer app rather than a research demo.
Meta compared the finished system against Runway Characters and HeyGen’s LiveAvatar, describing them as the two leading commercial avatar products, using each company’s own live-call experience rather than a lab benchmark. Raters said they preferred Muse Realtime Avatar 78 percent to 22 percent over Runway Characters and 88 percent to 12 percent over HeyGen LiveAvatar, across visual quality, lip sync, and character consistency. Meta’s own results carve out one exception: on mannerisms specifically, the company describes the Runway comparison as not statistically distinguishable from parity, a tie inside an otherwise clean sweep. Meta ran this comparison itself, and it did not publish who the raters were or how many conversations each one judged.
On Meta’s own hardware, a single session streams 448 by 768 pixel video at 25 frames per second with roughly 870 milliseconds between the end of a person’s turn and the first byte of the avatar’s response. Meta says serving optimizations, including four bit quantization and custom NVIDIA kernels built with NVIDIA’s help, let one GB200 chip handle 12 concurrent conversations, eight times more than an earlier, unoptimized version of the same model.
That efficiency gain is the real story for anyone pricing out a live avatar product. Runway and HeyGen both charge for character minutes today, and the cost of a live avatar session is set almost entirely by how many chunks of video a single GPU can serve at once. If Meta’s 60 times distillation and 8 times serving gain hold up outside its own infrastructure, it undercuts the unit economics rivals are currently selling against, not just the demo quality.
Meta says the technology carries an invisible watermark, Meta Video Seal, embedded in every frame without adding latency, and that the Muse app restricts avatar features to users 18 and older. The company also notes that examples in its announcement demonstrate what the underlying model can do and do not all reflect features currently live in the Muse app itself, a distinction worth remembering before assuming everything shown ships immediately.
Any team building or buying live-avatar tools should ask Runway and HeyGen for their own head-to-head numbers against Muse Realtime Avatar before renewing a contract, since the only comparison on record right now is the one Meta ran on itself.
Reported by Meta AI Research in a blog post published September 23, 2026.
