xg-chu/GAGAvatar

What it solves

GAGAvatar addresses the challenge of creating controllable 3D head avatars from a single image. It allows for one-shot 3D reconstruction of a person's head and enables real-time reenactment, meaning the avatar can be driven by another person's movements or expressions.

How it works

The project uses 3D Gaussian Splatting (3DGS) to reconstruct the head. It modifies the standard 3DGS renderer to allow Gaussians to carry 32-dimensional features, enabling more flexible and generalizable reconstruction. The system can be driven by a single image of a target person and a driving source (which can be an image or a video), using a tracking system to map movements from the driver to the avatar.

Who it’s for

This tool is designed for researchers and developers working in 3D vision, digital humans, and real-time animation, as well as those interested in in-person reenactment technology.

Highlights

• One-shot reconstruction: Creates a 3D head avatar from just one image.
• Real-time performance: Supports high-speed reenactment of the head avatar.
• 3DGS-based: Leverages Gaussian Splatting for efficient rendering.
• Controllable: Can be driven by images or videos to animate the head avatar.