How the attractiveness score works

Every formula, target range and weight behind the facial balance score, and what the test deliberately leaves out.

Last updated: 2026-11-18

This page documents exactly how the attractiveness test turns a photo into a facial balance score, including its limits. Nothing is hidden behind a "secret algorithm".

Step 1: finding the face

The test uses Google's MediaPipe Face Landmarker, an open-source model that runs inside your browser. It places 478 points on the face, including the centre of each iris. The model files are served from this site and cached by your browser; your photo is never sent to us or anyone else.

If there is more than one face in the photo, the largest is measured. The model also reports facial expression signals, which the test uses only to warn you about a smile, an open mouth or closed eyes.

Step 2: straightening the head

Before measuring, the points are rotated into a face-aligned frame: the horizontal axis runs between the two sides of the face, the vertical axis from the chin to the top of the forehead, and the origin sits between the eyes. This removes head tilt. A head turned sideways can't be fully corrected from one photo, so the test warns you when it detects a turn of more than about 12 degrees or a nod of more than 15.

Step 3: the five scored measurements

Each measurement gets a sub-score from 3 to 10. It scores 10 at the ideal value and falls along a bell curve toward 3 as the value moves away; the spread sets how fast. There is no flat top, so a value just inside the usual range scores less than one right at the ideal. The usual range is shown on each result card for reference.

MeasurementHow it is calculatedIdeal (spread)Weight
Symmetry45 left/right point pairs mirrored across the midline; average mismatch as % of face width, shown as 100 − 10 × mismatchmismatch 0% (2%)3
Middle vs lower third(brow to nose base) ÷ (nose base to chin)1.00 (0.13)1.5
Lip position(nose base to lip line) ÷ (lip line to chin)0.52 (0.08)1
Eye spacinginner-corner gap ÷ average eye width1.20 (0.12)1.5
Eye tiltaverage angle from inner to outer eye corner+4° (4.5°)0.75

The weighted average of the five sub-scores (3 to 10) is turned into the score you see, from 2 to 9.8. Averages below 4 count as 2; above that the scale curves, so each step near the top takes more than the one before. On our front-facing test photos a typical result is about 6 to 7, and only photos close to ideal on every measurement reach 9. A perfect 10 isn't on the scale, because no photo measures a face perfectly.

Why these ranges

The thirds and lip position follow the proportions taught in art and facial anatomy. The eye-spacing ideal is higher than the textbook "one eye width" because the model's corner points sit slightly inside the visible corners; it was set from what the model measures on front-facing test photos. The eye-tilt range follows the common finding that most eyes tilt slightly upward.

Why symmetry counts the most

Of all the proportions, symmetry has the most consistent link with attractiveness ratings in face research, and it is also the one the model measures most steadily: in our tests the same face, mirrored or tilted by 12 degrees, gave nearly the same symmetry reading.

Measured but not scored

Nose width, mouth width, upper-to-lower lip ratio, face width-to-height, face length-to-width and jaw width are shown under "More proportions". They don't count toward the score because a smile, an open mouth or the distance between the camera and the face changes them by more than the differences between faces.

What we deliberately don't do

  • No golden-ratio mask. The idea that beautiful faces follow 1.618 is popular but not supported by face research, so the test doesn't use it.
  • No "trained on ratings" AI score. Models trained to copy people's attractiveness ratings repeat the biases of whoever did the rating, and can't explain their answer. Every number here can be checked.
  • No comparison with other users. Photos are never collected, so there is no league table, and the score is not a percentile.

Limits you should know

  • Lenses distort. A phone held close to the face enlarges the nose and narrows the face. Distance matters more than the phone.
  • Light and hair move the points. Shadows and hair over the jaw or brows shift landmarks by a few pixels, which can move a ratio by a few hundredths.
  • One photo is one sample. The same person can score half a point apart in two photos. Take two or three and look at the average.
  • Proportions are not attractiveness. Skin, expression, grooming, style and personality matter a great deal and aren't measured.