A real-time avatar is an AI Avatar that is rendered live while the conversation happens, not recorded in advance. It listens and answers in speech, with a face that moves in sync, so a reply feels like a video call rather than a played-back clip.
Each turn of the conversation passes through four layers: speech recognition turns the person's words into text, a language model (LLM) decides what to say, text-to-speech turns the answer into audio, and the avatar model turns that audio into synchronised video. The video is streamed to the browser or app, usually over WebRTC.
Every layer adds delay, so vendors quote different latencies. Avatar render latency covers only the face; end-to-end latency covers the full reply.
Genesis 2.0 renders the Avatar in about 100 ms and streams 1080p at 35 FPS. With Managed Agents, the full reply arrives in about 1.0–1.2 s. Because Genesis 2.0 renders a learned 3D model of the person rather than generating every frame with a large diffusion model, one entry-level NVIDIA GPU runs 12 live 1080p Avatars. See how the AI agent stack fits together.
No. Every Beyond Presence Avatar is rendered live during the conversation, so it answers what each person actually asks.
Genesis 2.0 renders the Avatar in about 100 ms. With Managed Agents, a full reply starts about 1.0–1.2 s after the person stops speaking.
Yes. The Speech-to-Video API adds the face and keeps your LLM, voice and orchestration as they are. LiveKit is the recommended integration.