invisible layer / face inference
experiment · face inference
your face
is a data stream.
Point your camera at your face. A model running entirely inside your browser — no upload, no server, no account — will infer your emotional state, estimate your age and gender, map 68 biometric landmarks, and derive an identity hash from your facial geometry. This experiment transmits nothing. Real-world deployments do not make that promise.
Passive face inference is already standard infrastructure. Airport kiosks, supermarket self-checkouts, digital billboards, ATMs, and smart doorbells routinely estimate age, gender, attention, and emotional state — without disclosure. In the EU this is technically regulated. In practice it runs everywhere, in silence, at 30 frames per second.
▶ what is the model actually measuring?

Three neural networks run in sequence on every frame:

  • Face detector — locates bounding boxes at multiple scales. Sub-30ms on modern hardware.
  • Landmark network — maps 68 anatomical keypoints: eye corners, nostril base, lip edges, jawline. Sub-millimetre resolution relative to face size.
  • Expression classifier — reads muscle activation patterns from landmark geometry. 7 classes: neutral, happy, sad, angry, fearful, disgusted, surprised.
  • Age / gender estimator — regression over tens of thousands of labeled images. Age ±5 years; gender binary probability.

The geometry hash is derived from expression-stable ratios: inter-ocular distance, face width/height, nose width, mouth width. These change slowly, if at all, over years. They are the basis of face recognition systems at scale.

Everything runs client-side. Nothing leaves your device. That is the exception, not the rule.