How the FaceCheckup face test works, and how accurate it is

Your browser runs Google’s MediaPipe face model on your photo, places 478 points, checks the photo is usable, corrects for head turn, and measures distances, ratios and angles. Each number is compared with a published population range where one exists. It is shown with a ± range from re-measuring three re-framed copies. Nothing is uploaded.

1. Finding the face: 478 landmarks, on your device

The scan uses the MediaPipe Face Landmarker, an open-source model from Google. It returns 478 three-dimensional points: 468 across the face, plus 10 on the irises that locate the pupils. It also returns 52 expression scores and a matrix describing the head’s pose. We serve the model and its WebAssembly runtime from our own domain, and they run in your browser tab. That is why your photo never needs to leave your device, and why the first scan takes a few seconds while the model loads.

A second small model, MediaPipe’s hair segmenter, marks which pixels are hair. We use it only to find the hairline on the midline of the forehead, because the face mesh stops short of it. When hair covers the forehead, or the hairline isn’t clear, measurements that need it (the upper third, face length, eyes-to-mouth ratio) are hidden rather than guessed.

2. The photo-quality check

Most “random” results from face-rating apps come from the photo, not the face. Before scoring, we check:

• Head turn (yaw): warns above 8°, asks for a retake above 15°.
• Chin up or down (pitch): warns above 10°, retake above 18°.
• Head tilt (roll): corrected automatically; warns above 12°.
• Resolution: warns below 60 pixels between the pupils; retake below 35.
• Close-up framing: warns when the face fills more than 80% of the frame. Close lenses enlarge the nose and narrow the face (see the camera-distance research in section 7).
• Expression: a smile or open mouth changes jaw, cheek and eye geometry, so neutral is best.
• Eyes: closed or squinting eyes are flagged.
• Light: photos that are too dark or overexposed, or lit from one side, are flagged.

3. Pose correction

Even a 10° head turn makes one half of the face look wider and reads as asymmetry. So before measuring, we use the head-pose matrix to rotate the 3D landmarks back to a straight-on view, then take the 2D measurements from that. On one of our test selfies, taken with the head turned about 10°, uncorrected asymmetry measured 4.70% of face width. After correction it measured 1.06%, and the two half-widths went from a 72% match to a 95% match. The overlay you see is still drawn on your original photo, so the lines sit where your features really are.

4. The ± range

Every number is measured four times: on your photo, and on three copies that are slightly cropped and shifted. Small changes like these move a landmark by a pixel or two. The ± shown is half the spread between the four results, with a minimum so we never claim more precision than the method has (for example ±0.4° for canthal tilt and ±0.1 percentage points for asymmetry). On the same test photo, canthal tilt measured 7.01°, 7.81°, 7.31° and 7.46° across the four frames, so it was reported as about 7.4° ± 0.4°. Adding more photos with “+ Add a photo to average” averages across all of them, and usually narrows the range further.

5. The Balance Score formula

The Balance Score runs from 0 to 10 and measures two things research links to how faces are perceived: left-right symmetry, and how close proportions sit to published population averages. It starts at 10 and loses up to 6 points for asymmetry and up to 4 for proportions:

Why a free first 0.8%: no real face is perfectly symmetric, and landmark placement has a little noise of its own. MediaPipe’s published landmark error is about 2–4% of the distance between the eyes (Google’s model card), so small mismatches shouldn’t cost points. Symmetry carries most of the weight because it is the factor the research supports most consistently: a meta-analytic correlation with attractiveness of about r = 0.25 (Rhodes, 2006, as summarised by Jones & Jaeger, 2019). That is a real but modest link, which is one more reason the score is not a beauty verdict.

Why only two proportions: many “ideal ratios” circulate online, but few have been tested against how people actually rate faces, using landmarks we can reproduce from a photo. Pallett, Link and Lee (2010) did exactly that for these two, and the golden ratio failed similar tests repeatedly. Their faces were female, so for men these two ratios are a weaker guide. Every other measurement is shown for information and doesn’t change the score. If the hairline isn’t visible, the eyes-to-mouth ratio can’t be measured and costs nothing.

What the score is not. It doesn’t predict how attractive other people will find you. Attractiveness judgements depend on skin, expression, hair, grooming, culture and the person judging, and geometry explains only part of them. We call it a Balance Score so it is never mistaken for a beauty rating.

6. How the ranking works

Next to your score, the result says what share of a reference set of faces scored lower than you, for example “more balanced than 72%”. A score on its own is hard to read, and a percentile answers the obvious next question: is that high?

The reference set. We ran the same scan over the official portraits of the members of the 119th US Congress, as published by the @unitedstates project. Official congressional portraits are works of the US federal government, so they are in the public domain. Of 524 portraits, 10 failed the photo check, leaving 514 scores (read 2026-10-24). The median Balance Score was 8.4, a quarter of faces scored below 7.3, and a quarter scored above 9.2. We store only the scores, not the photos.

How it’s calculated. Percentile = (reference faces scoring below you + half of those scoring exactly the same) ÷ 514, rounded to a whole number.

Its limits. These are mostly middle-aged and older adults photographed in studio light, not a random sample of the public or of your country, so the percentile is a rough guide rather than a census. It ranks balance, which is symmetry and two proportions, not attractiveness. No public dataset ranks how attractive people are on a scale we could honestly compare you with, so we don’t claim one. We re-run the reference whenever the score formula changes.

7. Published values we compare against

Every other measurement is shown as a number without a judgement. Published norms do exist for several of them (eye spacing, nose width, facial thirds, the face width-to-height ratio), but they were measured with calipers or with differently placed landmarks, so comparing them with a MediaPipe measurement would mislead you. The guides discuss those studies in context. One example: the classical “equal thirds” rule held for 0% of a southern Chinese sample (Jayaratne et al., 2012), and the “one eye-width apart” rule for between 0% and 33% of people depending on the population.

8. Face shape

Stylists define face shapes by comparing forehead, cheekbone and jaw width, face length against width, and how angular the jaw is. Our detector measures those same things and finds the closest of seven shapes: oval, round, square, oblong, heart, diamond and triangle. It reports the top two with percentages. The reference points for each shape start from the proportions of Google’s canonical face model, and shift in the direction each stylist definition describes. Face shape is a convention rather than an anatomical category, so a 55% / 40% split simply means your face sits between two shapes. Our face shapes guide shows how to use either result.

9. Known limits

• Camera distance. A phone held close distorts the face, so for anything that matters, use a photo taken from further away with the zoom.
• A photo is 2D. Jaw angle, chin projection and cheekbone prominence are better judged in profile. We measure what a front view can show, and say so.
• Landmark accuracy. MediaPipe’s points are very consistent but not anthropometrically exact. Its “eye corner” may sit a little inside the true corner, for example. That is why we compare against published ranges only where the definitions line up.
• Glasses, beards, fringes and makeup can hide or shift landmarks. Take glasses off and pull hair back for the most accurate scan.
• Population differences. Published norms vary by sex, age and ancestry. Where a source gives one population’s range, we say which population it is.

10. Privacy

All of the above happens in your browser. No photo, landmark or measurement is sent to our servers or to anyone else, and nothing is stored after you close the tab. The details are in the privacy policy.

Ready to try it? Take the free attractiveness test, or start with a single measurement such as canthal tilt or facial symmetry.

Sukie · Founder and editor

Sukie built FaceCheckup because she always wanted a face tool that explains its numbers and tells you what to do with them, and nothing on the market did both. She checks every measurement on the site against published research and labels every tip for what it is: research, or stylist convention.

About this site · How pages are checked · Last updated October 24, 2026