Skip to content

Ultralight-Digital-Human

Run a lightweight digital human model on mobile in real time

Visit official site

What Ultralight-Digital-Human is

The model is described as an ultra-lightweight digital human model that runs in real time on mobile devices, with a design suited to mobile deployment. The model uses deep learning for face synthesis and voice simulation and supports audio feature extractors such as wenet and hubert.

What you can do with it

  • Run the model in real time on mobile devices with a lightweight design
  • Use the open-source code and contribute improvements
  • Train a personal digital human model following the detailed training steps
  • Apply streaming inference in real-time scenarios

Application scenarios

  • Create a virtual anchor to deliver news broadcasts
  • Build a virtual teacher for online classes in education
  • Produce a virtual idol for entertainment performances

Getting started

  • Install the required libraries and environment such as PyTorch and other dependencies
  • Download the wenet encoder.onnx file and place it in the specified directory
  • Prepare video and audio data and preprocess them
  • Train syncnet for better results

Pros & cons

✓ What we like

  • Open-source code invites community contributions and improvements
  • Streaming inference supports real-time scenarios

! What to watch out for

  • Setup requires PyTorch and other dependencies
  • Training calls for prepared video and audio data plus syncnet training

FAQ

Where does it run?

The material describes real-time operation on mobile devices.

Is the code open?

The material states all code is open source, with detailed training steps for a personal digital human model.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review