# Open avatar and face-tracking notices

## Avatar model

- File: `research-host/model.vrm`
- Source: vrm-c/vrm-specification `VRM1_Constraint_Twist_Sample.vrm`
- Fixed source revision: `821c11b250d8c70d5804ee13431e42bee56ea9c0`
- Author in embedded VRM metadata: pixiv Inc.
- License URL in embedded VRM metadata: https://vrm.dev/licenses/1.0/
- Embedded permissions: everyone may use the avatar; corporate commercial use,
  redistribution, and modified redistribution are permitted; credit is not required.
- Embedded restrictions: excessively violent, excessively sexual, political or
  religious, and antisocial or hate uses are not permitted.

The model is redistributed unchanged. Its SHA-256 digest is
`12c2b97e95e700783a6a550dc0eee2d7880aeedccef9ae67bc4c5a2f0f2631a2`.

## Browser runtimes

- `@pixiv/three-vrm` 3.5.5 — MIT License, copyright pixiv Inc. and contributors;
  full text: `/vendor/licenses/three-vrm-MIT.txt`.
- `three` 0.185.1 — MIT License, copyright 2010-2026 three.js authors;
  full text: `/vendor/licenses/three-MIT.txt`.
- `@mediapipe/tasks-vision` 1.0.1 — Apache License 2.0, MediaPipe Authors;
  full text: `/vendor/licenses/mediapipe-APACHE-2.0.txt`.
- Face Landmarker bundle source:
  `https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/latest/face_landmarker.task`.
  Its SHA-256 digest is
  `64184e229b263107bc2b804c6625db1341ff2bb731874b0bcc2fe6544e0bc9ff`.
  The component FaceMesh and Blendshape model cards identify Apache License 2.0.

Face landmark inference runs in the broadcaster's browser. The raw local camera
track is not added to the published stream; only the rendered avatar canvas is
published. MediaPipe states that input images stay on device, while API
performance and utilization metrics may be sent to Google; the host UI discloses
this before the broadcaster enables face tracking.