"AI avatar" covers two different technologies. Both show a face that talks, but they are built and paid for in opposite ways. If you are choosing a talking host for a website or a screen, this difference matters more than any feature list.
Video avatars: a face rendered on someone else's servers
A video avatar is generated as video. The provider's GPUs produce frames of a person speaking, usually from a filmed or photographed actor, and stream them to the visitor like a video call.
Strengths
- Photorealism. The face looks like a filmed person, because it is derived from one.
- Nothing to render on the visitor's side. Any device that can play a video can show it.
Trade-offs
- You pay for every minute. Each live conversation keeps a GPU busy on the provider's side, so pricing is per minute. As published in November 2026, Tavus plans ran from $22 to $975 a month for 60 to 4,000 minutes, with extra minutes at $0.26–0.35. D-ID rounds minutes up to 15 seconds and lets unused minutes expire each month.
- The face is a stock or filmed person. It can be custom, but it is a person, not a character designed for your brand.
- Streaming adds delay. The video has to be generated and sent before the visitor sees the answer.
3D avatars: a character drawn by the visitor's device
A 3D avatar is a model: a mesh with a skeleton and facial blend shapes, like a game character. The visitor's browser draws it with WebGL, and the conversation sends only text, audio and animation data, never video.
Strengths
- Cost doesn't grow with minutes watched. The visitor's device does the drawing; you pay for the AI and the voice, not for a GPU per conversation. That allows a fixed monthly fee.
- Your own character. Face, outfit, colours and voice can be designed for your brand.
- It reacts instantly. Expressions, eye contact and gestures are animated on the device, so the character is never frozen while the answer is prepared.
- It runs where your visitors are: phones, laptops and in-store kiosks, from one web page.
Trade-offs
- Stylised rather than filmed. A browser-rendered character looks very good, but not like video of a real person.
- It has to run on ordinary phones, which takes careful optimisation: model size, textures and fallbacks for weak devices.
The third option: cinematic 3D streamed from Unreal Engine
There is a middle path for flagship screens. A MetaHuman-quality character is rendered in Unreal Engine on a cloud GPU and streamed to the screen (Pixel Streaming). It looks cinematic and runs on any display, but like a video avatar, every live viewer occupies part of a GPU, so it suits a lobby, a showroom or an event stand rather than every website visitor.
Side by side
| Video avatar | 3D in the browser | 3D streamed from Unreal | |
|---|---|---|---|
| Rendered by | Provider's GPUs | Visitor's device | Your cloud GPU |
| Looks | Filmed person | Designed character | Cinematic character |
| Billing | Per minute | Fixed monthly fee | Per GPU hour |
| Your own character | Limited | Yes | Yes |
| Best for | Recorded videos, short sessions | Every website visitor, phones | Flagship screens and events |
Which one should you pick?
- Many visitors, long conversations, a brand character: 3D in the browser. Costs stay flat as traffic grows.
- A few high-impact screens: cinematic Unreal streaming.
- Pre-recorded explainer videos rather than live conversation: a video avatar platform is the simplest tool.
Lobbyface builds the first two: browser 3D hosts for websites and screens, and Unreal-rendered characters for flagship spaces, on the same AI and voice. See how the packages compare, or read what an AI avatar does that a chatbot doesn't.