OdinVTube is a lightweight, open-source avatar runtime for desktop.
It is written in Odin and rendered with raylib.

The goal is a small, understandable stack you can run fully offline: track motion into parameters, map those to an avatar, and stream or capture the result (for example with OBS chroma key).

Repository: github.com/aquaticcalf/odin-vtube

• Desktop window for a 2D avatar (procedural layers or optional PNG)
• Tracking inputs → smoothed model parameters (head, eyes, mouth, brows, breath)
• Mouse, idle, and demo tracking modes; keyboard for expression control
• Simple spring-style secondary motion
• Built-in expressions and hotkeys
• Optional green-screen background for OBS
• Local plugin API on 127.0.0.1 (JSON over TCP)
• Optional personal face model training pipeline (MediaPipe teacher → small ONNX student)

It is intentionally minimal. Full Live2D Cubism mesh support is not included yet; that would use the official Live2D Cubism SDK under Live2D’s terms.

You need the Odin compiler with the bundled vendor:raylib.

cd odin-vtube
build.bat

Or:

odin build . -out:odin-vtube.exe

Run from this directory so assets/ and configs/ resolve:

odin-vtube.exe

Edit configs/default.json:

OdinVTube uses its own layered JSON format (e.g. assets/models/default/avatar.json).
No Live2D Cubism SDK and no paid license for the format itself.

1. Run the app, press F5
2. Drop PNGs from Krita / Paint (or use Import folder)
3. Or Add part for procedural shapes (head, eyes, mouth, hair, …)
4. Drag the blue handle or use arrow keys
5. Bind head / body motion presets
6. Ctrl+S or Save model
7. F5 again to return to stream view

Ideal art path: draw layers in Krita → export each as PNG → drop into F5.
Simple path: one full character PNG from MS Paint → drop as billboard.

Full guide: docs/MODELING.md.

• Procedural layer kinds: head, body, hair, eye_l, eye_r, mouth, brow_l, brow_r, blush
• Optional image for a full-character PNG (billboard)
• Layer textures live under textures/ next to avatar.json
• Layers bind to params like ParamAngleX, ParamMouthOpenY, etc.

Press F9, then in OBS: Window Capture → window OdinVTube OBS → filter Chroma Key (Green).

Full steps: docs/OBS.md.

The server listens only on 127.0.0.1 (never on the public network by default).

Protocol: one JSON object per line over TCP (default port 8001).

{
  "apiName": "OdinVTubeLocalAPI",
  "apiVersion": "1.0",
  "requestID": "1",
  "messageType": "APIStateRequest",
  "data": {}
}

Supported message types include:

• APIStateRequest
• StatisticsRequest
• AuthenticationRequest / AuthenticationTokenRequest
• InputParameterListRequest
• Live2DParameterListRequest
• InjectParameterDataRequest
• CurrentModelRequest
• HotkeyListRequest

Example parameter inject:

{
  "apiName": "OdinVTubeLocalAPI",
  "apiVersion": "1.0",
  "requestID": "2",
  "messageType": "InjectParameterDataRequest",
  "data": {
    "parameterValues": [
      { "id": "MouthOpen", "value": 1.0, "weight": 1.0 }
    ]
  }
}

Train a small face model on your own footage (optional).

• Full guide (including exact video checklist and durations):
  personal_face/README.md
• Colab notebook:
  personal_face/notebooks/Personal_Face_Train_Colab.ipynb

Film yourself doing all of these (see the full table in personal_face/README.md):

1. Neutral face (~1–2 min)
2. Talking while looking at cam (~5–8 min) — most important
3. Mouth open/close (~1–2 min)
4. Smile / laugh (~1–2 min)
5. Blinks + hold eyes closed (~1–2 min)
6. Look left / right (~1–2 min)
7. Look up / down (~1–2 min)
8. Head tilt / roll both ways (~1–2 min)
9. Talk while turning head (~2–3 min)
10. Raise brows / slight frown (~1–2 min)
11. Lean closer / farther (~1 min)
12. Optional: ¾ side views, dimmer lighting

Use normal desk lighting and the same glasses setup you will use live.

Users do not need OBS. Open the guided recorder:

• Folder: personal_face/recorder/
• Windows: double-click personal_face/recorder/open_recorder.bat
• Or open personal_face/recorder/index.html in Chrome / Edge

It shows each step on screen, records your webcam, and lets you download the file for Drive / Colab.

The Colab notebook clones this repo, labels video with MediaPipe, trains a compact student network, and exports ONNX to Google Drive.

odin-vtube/
  main.odin, *.odin     # runtime
  assets/               # default avatar
  configs/              # app settings
  personal_face/        # training pipeline (Python + Colab)
  build.bat

This project is licensed under the MIT License.

Third-party libraries (raylib, MediaPipe, OpenCV, Live2D Cubism if added later, etc.) keep their own licenses.