How it works

The four steps

1. Your browser reads the file. The image never leaves the device: it is decoded locally, drawn on a canvas and analysed in memory.
2. 478 landmarks. A face-landmark model (MediaPipe Face Landmarker) returns 478 points of the face, including a 36-point outline of the jaw, chin and forehead.
3. Four proportions. From that outline we measure the length of the face and its width at the forehead, at the widest point (the cheekbone/temple line) and at the jaw, plus the width of the chin.
4. Six profiles. Each shape is an ideal combination of those proportions. The face is scored against all six and the percentages show how close it sits to each one.

Why the widest point is searched, not assumed

The widest part of a real face is usually around 30 % down from the top of the outline, not at mid-height. Fixing the measurement at 50 % makes most faces look as if the forehead were wider than the cheekbones, and pushes them towards "heart". We search between 10 % and 60 % and take the true maximum.

Why you also get a second shape

Face shape is a continuum, not a set of boxes. When the top two shapes are close, the result is labelled a low or medium match and the second option is shown. That is more useful — and more honest — than a confident single answer.

Photo quality checks

• Exactly one face in the frame.
• The face fills at least a quarter of the photo height.
• The photo is at least 512 px on the short side.
• The head is level and facing the camera.
• Enough light: a very dark photo can't be measured reliably.

If any of these fail, the tool explains what to change instead of guessing.

What it does not do

It does not identify people, estimate age, gender or ethnicity, analyse skin, or store images. It is a styling aid, not a medical or cosmetic diagnosis.