Tavus — AI Video Generator

Phoenix 4 Unifies emotional state control, active listening behavior, and continuous facial motion in one real-time system. Runs full-duplex, meaning it listens and responds simultaneously instead of taking turns. Generates behavior continuously with millisecond-level latency for live conversation responsiveness. Expands explicit control over how emotion is expressed during both speaking and listening states. Trains on large-scale human conversation data to learn realistic non-verbal conversational behavior at runtime.

Phoenix 3 Moves from talking-head style emphasis to full-face rendering, animating eyebrows, cheeks, eyes, and mouth, not just lips. Adds real-time dynamic emotion control so expressions adapt to conversational context or explicit settings. Targets hyper-realistic micro-expressions to push conversations closer to human-like presence. Shifts the product framing toward real-time conversational video interaction, not just avatar video generation. Becomes a core pillar in a multi-model conversational stack designed to “see, listen, and engage” in real time.

Phoenix 2 Switches the technical direction toward 3D Gaussian Splatting style 3D scene representations for avatar rendering. Improves controllability by leaning on 3D intermediates instead of purely 2D GAN style pipelines. Targets better temporal consistency for longer sequences by moving away from classic 2D talking-head limits. Prioritizes fast rendering style workflows (3D rendering pipeline focus) for practical avatar applications. Frames the model evolution explicitly as NeRF to 3DGS, with 3DGS presented as the preferred path forward.

Phoenix 1 Trains a lifelike digital replica from a short 1 to 2 minute video. Generates text-to-video from new scripts after the replica is trained. Produces full head and shoulders performance with natural expression and emotion. Uses a pipeline that includes TTS, 3D reconstruction, script-driven facial animation, and high-fidelity rendering. Handles wide variation in voices, appearances, and real-world capture conditions.