A Pretty Scale Test That Shows Its Work

PrettyScale.com is the best-known pretty scale and the brand behind most searches for this phrase. Its own help page explains that it 'measures the face beauty by studying and testing the placement and sizes of different face features', that results are 'based on the face shape, distance between eyes and lips, mouth size and face symmetry', and that 'sometimes golden ratio is also used'. You place lines and dots on your photo following its instructions, and the help page adds that 'results can be different based on how you place lines and dots on your face'.

The score is a percentage. PrettyScale describes it with a joke unit, the 'micro Helen', after Helen of Troy, 'the face that launched a thousand ships': a score of 80% means 80 micro Helens. Its home page also carries two warnings worth taking seriously: 'Please do not start if you have low self-esteem or confidence issues', and results 'could be incorrect'.

Whatever the app, a score of this kind comes down to the same steps: locate a handful of points, measure the distances between them, compare each distance with a target ratio, and turn the gaps into a percentage. The targets are the weak point, because they come from art traditions and the golden ratio, not from measurements of real people.

Goshtasbi, Hakimi and Wong (2024) tested five face-rating websites, prettyscale.com among them, against a panel of 24 human raters. Everyone rated the same 40 frontal photos. The sites' average scores rose and fell with the human averages, a correlation of r = 0.84, so they ranked faces in a broadly similar order. But they were consistently generous: the sites averaged 6.9 out of 10 against 5.0 from the people, and 30 of the 40 faces (75.0%) were scored significantly higher by the sites than by humans.

Two limits matter. The 40 faces were AI-generated images of white women, a narrow set, and the paper's abstract doesn't report PrettyScale's results separately from the other four sites. So the study can't tell you how accurate PrettyScale is on your face. What it does show is that a confident number from a rating site can sit two points away from what people actually think, and in either direction you would have no way to tell from the screen.

The classical canons say a face should be five eye-widths across, with the gap between the eyes equal to one eye width and the nose as wide as that gap. Measured faces rarely fit. The orbital canon, eye gap equal to eye width within a millimetre or so, held in 0% of 103 southern Chinese adults (Jayaratne et al. 2012), 29% of Saudi men and 25% of Saudi women (Al-Sebaei 2015), and 33% of Kenyan men and 27% of Kenyan women (Virdi et al. 2019). A pooled analysis of raw data from 27 studies found that no ethnic or gender group consistently matched the canons or the golden ratio (Khoshab et al. 2022).

The golden ratio fares no better as a beauty formula. Burusapat and Lekdaeng (2019) measured Miss Universe and Miss Universe Thailand winners and found 12 of 16 facial golden ratios differed significantly from phi in both groups. Naini (2024) concluded there is 'no convincing evidence' that the golden ratio is linked to facial beauty. A pretty scale built on these targets can mark a face down for being an ordinary, healthy human face.

Research on attractiveness gives a more nuanced picture than any score. When many people rate the same faces, their averages agree closely, even across cultures (Langlois et al. 2000). But each person's own taste is large: Hönekopp (2006) found that private taste is about as powerful as shared taste. That means a face one person rates as average, another person may find striking, and both are being honest.

Things a pretty scale never measures make a big difference too. Bryan, Perona and Adolphs (2012) found that the same faces photographed at 45 cm were rated less attractive than at 135 cm. Close selfies also stretch central features: Ward and colleagues (2018) modelled the base of the nose appearing about 30% wider relative to face width at 12 inches than at 5 feet. Expression, lighting, skin, hair and the way you look at the camera all shift ratings, and none of them is in the formula.

So if a rating site gave you a low number, it has told you how your hand-placed dots compare with an old art formula, in one photo, on one day. It has not told you how pretty you are. If checking your looks has started to take up a lot of your time or make you feel low, the PSL scale page has a short note on getting support.

The scan finds 478 landmarks with Google's MediaPipe Face Landmarker, checks the photo first (head turn, chin angle, tilt, smile, closed eyes, resolution, lighting) and asks for a retake rather than scoring a bad picture. It then rotates your head pose back to frontal before measuring. On one of our own test selfies, with the head turned about 10° and a smile, the mirror mismatch fell from 4.70% of face width before correction to 1.06% after, which shows how much a slight turn can distort a symmetry reading.

Results open with your top three strengths, then the measurements, each drawn as a layer you can switch on: Outline, Symmetry, Thirds, Fifths, Eye tilt, Brow map, Widths, Jaw and Ratios. If you want to see your own rule-of-fifths spacing, switch on Fifths; you will see the numbers, not a pass or fail. Only three measurements are shown with a published reference beside them: pupil distance ÷ face width and eyes-to-mouth ÷ face length, compared with Pallett, Link and Lee's (2010) ratios of about 46% and 36%, and canthal tilt, shown with group averages for context.

The Balance Score sums those up from 0 to 10: up to 6 points for symmetry and up to 4 for how close the two Pallett ratios sit to the published values. It measures geometry. It is not a prettiness score, and we say that on the result screen. The golden ratio face page and the facial symmetry test page explain each part in more depth.

These come from FaceCheckup's sourced rulebook, each labelled with where the advice comes from. They change how you look in pictures, which is where most pretty scale scores start.