Free Attractiveness Test, No Sign-Up Needed

Everything below happens in your browser tab, in this order. The first scan takes a few seconds while the model loads.

Here's why the extra steps matter, using one of our own test selfies from October 2026. The head was turned about 10° and the person was smiling, and the quality check flagged both. We kept the photo on purpose, because it's the kind most people actually upload.

Before pose correction, the scanner measured 4.70% asymmetry, as a share of face width. After rotating the landmarks back to straight-on, it measured 1.06%. Most of the apparent asymmetry was the head turn, not the face. We then mirrored the same photo and ran it again: 3.63% before correction, 1.51% after. The left and right half-widths went from a 72% match before correction to a 95% match after.

The Balance Score runs from 0 to 10. It starts at 10 and loses points in two places only: symmetry (up to 6 points) and proportions (up to 4 points). The full formula, with every constant, is on the How it works page.

Symmetry. The scanner mirrors 220 pairs of landmarks across a midline through your nose and chin, and averages how far each point lands from its partner, as a percentage of face width. The first 0.8% is free, because no real face is perfectly symmetric and the landmarks carry a little noise of their own. After that, every further 0.6% costs one point, up to 6. On our test photo above, the corrected 1.06% costs under half a point. The uncorrected 4.70% would have cost all six.

Proportions. Only two ratios count, because they are the two that Pallett, Link and Lee (2010) tested against how people rated faces: the distance between the pupils at about 46% of face width, and the eyes-to-mouth distance at about 36% of face length. Within 2 percentage points of either value costs nothing. Beyond that, each further 2 points costs one point, up to 2 per ratio. If your hairline isn't visible, face length can't be measured, so that ratio is scored as neutral rather than guessed.

Everything else, including thirds, fifths, fWHR, midface ratio and jaw width, is shown as a number with no judgement. Published norms for those were measured with calipers or differently placed landmarks, so comparing them with a photo would mislead you. Canthal tilt is shown with context (published group averages run from about 4.7° to 8.8°) but isn't scored.

People's ratings of faces agree more than you might expect, but much less than a single score suggests. Both are true, and a fair test has to respect both.

Agreement is real at the group level. A meta-analysis by Langlois and colleagues (2000) found that the average of many raters is highly reliable: about r = .90 within a culture, .88 across ethnic groups and .94 across cultures. That describes a crowd's average, not whether two people agree. Hönekopp (2006) found that private taste is about as powerful as shared taste, so roughly half of what one person finds attractive is their own preference.

Symmetry matters, modestly. Rhodes (2006) concluded from a review and meta-analyses that symmetry, averageness and sex-typical features are all found attractive, in male and female faces and across cultures. The meta-analytic link between symmetry and attractiveness is about r = 0.25, as summarised by Jones and Jaeger (2019). In their own study of natural female faces, shape symmetry added little once other factors were counted. That's why symmetry carries most of our score's weight but the score still can't be read as a verdict.

Averageness is attractive. Langlois and Roggman (1990) found that composite faces, made by blending many photos, were judged more attractive than almost all the individual faces they were made from. That's why we compare only with published averages, never with a supposed ideal.

Many face-rating sites still grade you against the golden ratio, 1.618. We don't, because the evidence keeps saying it doesn't describe attractive faces. Kiekens and colleagues (2008) tested 19 supposed golden proportions: only 4 related to attractiveness ratings, together explaining 16% of the variance. Holland (2008) found the famous Marquardt “phi mask” best fits masculinized white fashion models, not faces in general.

The newer studies agree. In Hong Kong Chinese faces, 42 of 77 measurements in the attractive female faces differed significantly from the supposed ideal (Zwahlen et al., 2022). In Miss Universe and Miss Universe Thailand winners, 12 of 16 facial golden ratios differed significantly from phi (Burusapat and Lekdaeng, 2019). A 2024 review by Naini concluded there is no convincing evidence linking the golden ratio to facial beauty.

The classical canons fare no better. The “one eye-width between the eyes” rule held for between 0% of a southern Chinese sample and about a third of North American white participants, depending on the study (Jayaratne et al., 2012; Farkas et al., as reported by Al-Sebaei, 2015). Equal facial thirds were found in 0% of that southern Chinese sample. A raw-data analysis of 27 studies found that no ethnic or gender group consistently matched the ideals of either the canons or phi (Khoshab et al., 2022).

So the Thirds and Fifths layers show your own proportions without marking you down. The golden ratio face page explains what a phi grid can't tell you, and measures the two research-tested ratios instead.

If three tests gave you three different numbers, they were mostly measuring the photo, the camera and their own scale, not your face.

The scales are inflated. Goshtasbi, Hakimi and Wong (2024) ran 40 frontal face images through five AI rating websites, including PrettyScale and attractivenesstest.com, and compared them with 24 human raters. The sites ranked the faces in a similar order to the humans (r = 0.84), but scored them higher: an average of 6.9 ± 0.9 against the humans' 5.0 ± 1.8. In 30 of 40 images (75%), the AI score was significantly higher. A test that tells everyone they're a 7 is a pleasant number, not information.

The camera distance changes the face. Ward and colleagues (2018) modelled a selfie taken about 12 inches (30 cm) away and found the nose base appears roughly 30% wider relative to the face than from 5 feet (1.5 m). Pressler and colleagues (2022) measured nasal length 6.4% longer in 12-inch selfies than in clinical photos. Distance changes judgements too: Bryan, Perona and Adolphs (2012) found the same faces were rated less trustworthy, competent and attractive when photographed from 45 cm than from 135 cm.

Pose and expression change the geometry. As our own test photo showed, a 10° head turn more than quadrupled the measured asymmetry. A smile widens the mouth, lifts the cheeks and narrows the eyes.

Every landmark model has error. Google's model card for the face mesh we use reports average landmark error of about 2.1% to 3.9% of the distance between the eyes, depending on the mode and face region. Human annotators placing the same points by hand differed by 2.56%. That is why every number here carries a ± range.

Five minutes of setup does more for your result than anything else. Each step is labelled with where it comes from.

The scan also finds your face shape: oval, round, square, oblong, heart, diamond or triangle. These are stylist and optician conventions, not anatomy, so the detector compares the same things they do. That means forehead, cheekbone and jaw width, face length against width, and how angular the jaw is. An oval face, for example, is usually described as about one and a half times as long as it is wide, widest at the cheekbones, with a softly rounded jaw.

You get the top two shapes with percentages, because many faces sit between two. A 55% / 40% split isn't an error. It means advice for either shape will work, and you can pick whichever you like better. The face shape detector shows the widths it compared, so you can see why it chose what it did.

Your result ends with up to eight practical tips picked for your measurements, never a verdict on your worth. Every rule behind them has at least two published sources and an honest label: stylist, optician, makeup-artist or photographer convention, or research-supported. Examples:

The free test is the whole scan, not a teaser: no scan limit, no blurred results, no “unlock your score” and no trial that turns into a weekly charge, which is a recurring complaint in face-rating app reviews.

The Full Face Report is a one-time $12.99 payment, not a subscription. It averages 3 to 5 photos, ranks your photos for a profile picture, and gives your complete style plan and every measurement in one printable table. It runs on your device like the free scan, and the report page shows exactly what it contains before you buy. There are no refunds for change of mind. If the report fails to load, is missing anything promised, or you were charged twice, we fix it or refund you in full.

Face geometry is sensitive, so we built the test so we never receive it. The model runs inside your browser tab. No photo, landmark or measurement is sent to our servers or anyone else's, and nothing is kept after you close the tab. Before your first scan you tick a consent box. The free scan stores only that tick, in your browser's session storage, which is deleted when you close the tab. If you buy the report, a purchase token is also kept in your browser so the report stays unlocked.

The share card is made on your device too, and includes your photo only if you choose to add it. The test is for adults scanning their own photos, so please don't use it to rate or embarrass anyone else. The privacy policy has the full details.

FaceCheckup is built and edited by Sukie, the site's founder. She made it because none of the face tests she tried explained their numbers or said what to do with them. She isn't a doctor or a licensed stylist, and the site doesn't pretend otherwise. That's why every number on the site comes from a named, published source, and every styling tip is labelled as the convention it is. The full method is on the How it works page, and every study cited here is listed below.