Avatar Models

Two avatar models are available — for both companions and digital twins. Both produce real-time video avatars. The difference is the reference input and the resulting quality.

Set the version field to "v1" or "v2" when creating a companion or digital twin. If omitted, it defaults to "v1".

v1 — Image input

Provide a pictureUrl — a publicly reachable HTTPS URL of a reference image. For best results, use a 16:9 image with the person looking into the camera, framed waist up, and well-lit.

{
  "pictureUrl": "https://example.com/images/avatar.png"
}

v1 is fast and produces strong photorealism. It works well for quick iteration, internal testing, and use cases where speed matters more than pixel-perfect accuracy.

v2 — Video input

Provide a videoUrl — a publicly reachable HTTPS URL of a reference video showing the person speaking.

{
  "videoUrl": "https://example.com/videos/person-speaking.mp4",
  "version": "v2"
}

Why v2 produces better results

A still image gives the system a single frame to work from — it has to guess how the person's face moves when speaking. A video gives it the real thing. The system learns the person's unique mouth movements, facial micro-expressions, head motion patterns, and how their features shift during natural speech. This means:

• Lip sync stays accurate even during fast speech, complex words, and language switching
• Expressions feel natural — subtle brow raises, eye movements, and smiles that match the tone of what's being said
• Facial hair renders cleanly — beards, mustaches, and stubble that move naturally with the jaw instead of clipping or blurring

When to choose v2

Use v2 when the avatar represents someone your users will interact with directly — customer-facing support agents, executive avatars, brand representatives, sales leads, or any use case where the realism of the avatar affects trust and engagement.

Tips for a great reference video

The quality of the output depends heavily on the quality of the input. For the best v2 results:

• Duration: 1–5 minutes of continuous natural speech
• Content: The person should speak naturally, not read a script — conversational tone produces better results than stiff delivery
• Framing: Head and shoulders, centered, with some space above the head
• Lighting: Even, front-facing light. Avoid harsh shadows, backlighting, or overhead fluorescents
• Background: Clean and uncluttered. A plain wall or blurred background works best
• Camera: Eye-level, stable (tripod or resting on a surface). Avoid handheld footage
• Audio: Clear audio helps the system align lip movements — use a quiet room