Face Shape Data

Everything the face shape detector uses to turn a photo into one of seven shapes, published as plain files so anyone can check it, copy it or argue with it.

What is in here

How the numbers are produced

  1. A face landmark model finds 468 points on the photo. The model file is served from this site and runs on your own device; the photo is never uploaded.
  2. Four widths and lengths are read off those points: forehead width, cheekbone width, jaw width and face length.
  3. All four are expressed as a percentage of face length, which is fixed at 100. They are relative proportions of your own photo, not millimetres or inches.
  4. Three ratios follow: length ÷ cheekbone width, forehead width ÷ cheekbone width, and jaw width ÷ cheekbone width.
  5. Each of the seven target profiles is scored by how far those three ratios sit from it, scaled by the tolerances in the CSV. The closest profile wins; the score shown in the tool is that closeness measure, not a probability.

The seven target profiles

Shapelength ÷ cheekboneforehead ÷ cheekbonejaw ÷ cheekbone
Oval1.450.860.76
Round1.180.900.86
Square1.220.970.95
Heart1.421.000.70
Diamond1.500.740.72
Oblong1.750.880.84
Triangle1.320.801.02

A ratio may sit up to 0.30 (length ÷ cheekbone) or 0.12 (the two width ratios) away from a profile and still count as a match, which is why two shapes can score close together on the same face.

What this data cannot tell you

Using or citing this data

If you want to refer to these numbers, this wording is accurate: “Target ratio profiles and tolerances published by Face Shape Hub (faceshapehub.com/data/), retrieved [date].”

These files are published under the Creative Commons Attribution 4.0 International licence (CC BY 4.0): you may copy, redistribute and adapt them, including commercially, as long as you give appropriate credit — the citation line above is enough. The licence text is the authoritative version.