The system can animate a photographic portrait, full-body illustration, animal, or object from reference media during a live conversation.
Meta says the avatar model is an audio-driven Diffusion Transformer. It takes a stream of speech tokens, reference media, and recent generated video latents, then produces video in short causal chunks. The company says its distilled student model uses two unguided evaluations per chunk rather than the teacher’s 120 guided evaluations, a claimed 60x reduction.
For serving, Meta reports 448x768 portrait video at 25 fps and about 870 ms from the end of a user turn to the first byte of synchronized voice and video. It says a GB200 can support 12 concurrent real-time video sessions, an 8x capacity increase over its two-step BF16 baseline.
The source reports Meta’s own comparisons with Runway Characters and HeyGen LiveAvatar, but does not establish independent benchmark results. It also does not state public availability, pricing, supported platforms, or confirm a Muse desktop-app test.