How MogTracker measures a face

Every number the app produces comes from 478 facial landmarks detected in your browser. Nothing is uploaded, and no human reviews any result.

Landmark detection

A face landmark model runs locally in the browser via WebAssembly and returns 478 three-dimensional points across the face on every camera frame. Landmark positions are smoothed across frames before measurement, because raw per-frame output jitters by a pixel or two and that jitter is large enough to swing a ratio meaningfully. All processing happens on your device — the image never leaves it, and there is no server-side face pipeline to send it to.

Reference values

Each metric is scored against a reference derived from a reference face — Bomer for male scoring, Heard for female. Critically, those references were measured through the same landmark pipeline that scores users, not copied from anatomical literature. This matters because the mesh does not always place landmarks where a caliper would: the nostril landmarks, for instance, sit at the nasolabial fold and read about 5% wider than the true alar base. Referencing published anatomical figures against mesh-derived measurements would introduce that discrepancy into every score. Measuring both the same way cancels it.

Because of this, the reference values are a calibration rather than a claim about ideal anatomy, and changing a reference photo shifts every constant at once and rescores every user.

Scoring

Most metrics score on a Gaussian curve centred on the reference, so deviation in either direction reduces the score and the penalty grows with distance. Two exceptions — facial thirds and facial fifths — use linear scoring instead. A calibration run over 1,700 photographs showed the curve was too punishing for those two, returning near-zero scores for ordinary faces at modest deviations.

Seven metrics use different references for men and women, where a single unisex target would systematically misscore one sex. Four are scored identically for both. The composite is a weighted average with a harmony penalty applied when one metric falls far below the rest.

Why live camera only

Photo upload scoring was removed. Testing across a large photo set showed the same person scoring across an 87 to 96 point range depending on the photograph — lens, distance, angle and lighting moved the landmarks more than real facial differences did. A live feed lets the app enforce head pose limits before it will score a frame, which is the only way to make results comparable between people. Pitch, yaw and roll must all sit within a few degrees of neutral.

What this cannot tell you

The system measures geometric conformity to a fixed reference. It does not measure attractiveness. It cannot see skin, hair, expression, grooming, styling, movement or voice, and those account for a great deal of how faces are actually perceived. Two people with identical scores can look nothing alike and be received completely differently.