This project aims to provide a general framework that intermediate VTubers/programmers could use to build a custom VTuber avatar solution off of. It uses a Python backend for all of the face detection aspects, then pipes over the data to Unity for finetuning and displaying the avatar.
This means that you have the entire Unity pipeline at your disposal for VTubing and the simplicity of a Python backend as well. Currently, this project is powered by Google MediaPipe Face and requires only 1 RGB Webcam to use and no fancy GPU. Additional features may be added/developed upon request.
- Install Python and Unity (2021.3.17f1 was used, but any version close to that should be fine).
- pip install mediapipe
- Clone/download this repository.
- Run main.py using Python.
- Run the Unity project (SampleScene.scene)
- Your avatar will be automatically calibrated upon playing Unity.
- I do not have the rights to distribute the 3D model for demoing purposes. You will need to manually add it.
- For Programmers: Expect SWEEPING CHANGES to the architecture and code base going forward (currently not in a stable state). If you're just doing some experiments/learn this should be no problem, but I would not currently recommend using the present system for a long term project just yet.
If you are interested in commissioning an artist to create a custom vroid 3d model with 52 blendshapes, consider donating to my ko-fi (end product will be embedded in this project and open source for the community).
- See global_vars.py for some basic configuration options to speed up/improve precision of the detection.
- See AvatarFace script in the Unity inspector for configuration options (stylized behavior vs realistic, etc.)
- Good natural lighting goes a long way to aid face detection.
- This project aims primarily to create an appealing tracking result over an accurate one.