TypologySienna Chen
InputLightweight

Face tracking

The camera reads head turns and expressions.

The camera follows the face: head turns, mouth opening, eyebrows, blinks. Head-driven parallax, masks, characters that copy your expression. People are sensitive to their face being watched, so show a clear indicator.

Breaks when
lighting-dependentprivacy expectations
Also called
FaceLandmarker, face mesh, blendshapes, AR filter

Prompts

Head-driven parallax scene
MediaPipe FaceLandmarker with facial transformation matrices enabled: derive head yaw and pitch from the matrix, smooth with a One Euro filter, and move the Three.js camera on an arc (±0.25 rad yaw, ±0.15 rad pitch) around the scene so the screen behaves like a window. Blendshape jawOpen above 0.5 triggers a burst; eyeBlinkLeft/Right both above 0.6 pauses. Never render the video; only a small “tracking” dot.
Reactive character
Map 52 ARKit-style blendshape scores from FaceLandmarker onto a glTF avatar’s morph targets by name (lowercase match), scaled by a per-shape gain table, at 30 fps with 0.15 smoothing. Show which shapes are active in a side list. Explain the mismatch cases (missing targets) and how to log them once.
Mask that sticks
Attach a 3D mask mesh to the face using the transformation matrix (position, rotation, scale) with an occluder mesh of the head so the mask hides behind the nose and cheeks correctly. Match the camera’s field of view to the webcam’s to avoid drift.

Related

Resources · 1