This is a mirror of the two methodologies, not a verdict on a person. It reads everything it can straight from the image — geometry from 478 facial landmarks, plus skin texture, tone and under-eye from the pixels — so it’s drag-and-drop. But be honest about the ceiling: a transparent, on-device read is a
weak proxy. As
The Bone Pill shows, no honest instrument maps a face to a number, the features that decide most (and the whole profile) a front still can’t see, and the “shared score” dissolves the moment someone actually knows you. Pixel reads are rough and lighting-sensitive. When a trained model (
models/face-beauty.onnx) is present it scores the headline number — on-device, still no upload — and the geometry becomes the transparent breakdown beneath it; until then the heuristic scores both.
That model learned from SCUT-FBP5500 — a fixed panel of raters scoring one specific set of faces — so its scale carries their tastes, era and demographic skew, not a universal truth. Everything runs locally; the image never leaves your browser.
This is a transparent prototype, not a clinical tool — for a professionally-built facial assessment, see
QOVES.