Google Tests Gemini 3D Avatars for Android

https://www.findarticles.com/google-tests-gemini-3d-avatars-for-android/

Publish Date: 2026-02-04 08:06:00

Google appears to be preparing a new Gemini-powered feature that creates lifelike 3D avatars from a simple selfie video, hinting at a future where you can join a call without ever turning on your camera. Strings discovered in the latest Google app beta for Android reference a tool called “Characters,” a system that mirrors Apple’s Vision Pro Personas and Samsung’s Galaxy XR “Likeness” avatars.

What the beta hints reveal about Gemini Characters

In version 17.4.66 beta of the Google app, a Characters option surfaces in the Gemini launcher’s overflow menu. Tapping it brings up a flow to “record your likeness,” followed by a prompt to capture a short selfie video that the software uses to generate a 3D representation. Early testers report options to edit or delete previously recorded avatars, though the capture flow isn’t fully enabled yet.

One code string — “assistant_ui_account_wasabi_button” labeled as “Characters” — suggests this is meant to live alongside Gemini’s core UI. Based on placement and language, the feature looks intended for use in video calls or other live interactions, acting as a real-time stand-in when you’d rather not appear on camera.

How Gemini Might Build Avatars Without Headset Sensors

Apple’s Personas and Samsung’s Likeness leverage sophisticated sensors on headsets to capture depth and facial motion. Phones don’t have that same sensor array, so Gemini likely leans on AI reconstruction: estimating 3D geometry from 2D video using techniques such as monocular depth estimation, neural radiance fields, or Gaussian splatting, then driving a rigged face model with blendshape parameters.

Google has the building blocks. MediaPipe Face Mesh already tracks hundreds of facial landmarks from a standard camera. Google Research has published extensively on text-to-3D and view synthesis. Marrying those with Gemini’s multimodal capabilities could yield a pipeline that learns your appearance from a…

Source