What does a golden ratio face test actually measure?
A golden ratio face test compares distances on your face with values borrowed from classical art, usually 1.618. Most of them hand back one score. This one scores eight proportions separately and tells you which ones sit off target, in which direction, and by how much.
Run the golden ratio face test on a photo
Front facing, level, hair off the forehead, mouth closed.
Front-facing photo, even light, hair off the forehead
- Eight ratios, each with its own measured value, classical target and signed gap.
- Sorted worst first, so the answer to "which ones are off" is the top row.
- Two ratios depend on an estimated hairline and are flagged, because across our 43 measured faces they are wrong in the same direction every time.
- Runs in your browser. No upload, no account, no attractiveness score.
What each of the eight proportions compares
Three of these use 1.618. The rest come from the same classical canon but ask for two measurements to match, or for the face to be five eye widths wide. Calling all of them golden is a habit of the internet, not of the canon, so the target is written next to each one.
Face length to face width
hairline to chin, divided by cheekbone to cheekbone
Target 1.618. Across our 43 measured faces the median miss was 15.9% and 0% of faces landed within five percent.
Mouth width to nose width
mouth corner to mouth corner, divided by nostril to nostril
Target 1.618. Across our 43 measured faces the median miss was 4.4% and 56% of faces landed within five percent.
Chin to lips, against lips to nose
chin to the lip line, divided by the lip line to the base of the nose
Target 2.000. Across our 43 measured faces the median miss was 7.4% and 40% of faces landed within five percent.
Lower third to middle third
nose base to chin, divided by brow line to nose base
Target 1.000. Across our 43 measured faces the median miss was 6.2% and 42% of faces landed within five percent.
Upper third to middle third
hairline to brow line, divided by brow line to nose base
Target 1.000. Across our 43 measured faces the median miss was 43.5% and 0% of faces landed within five percent.
Face width in eye widths
cheekbone to cheekbone, divided by the width of one eye
Target 5.000. Across our 43 measured faces the median miss was 2.3% and 77% of faces landed within five percent.
Eye gap to eye width
the gap between the inner eye corners, divided by the width of one eye
Target 1.000. Across our 43 measured faces the median miss was 20.9% and 0% of faces landed within five percent.
Nose width to eye gap
nostril to nostril, divided by the gap between the inner eye corners
Target 1.000. Across our 43 measured faces the median miss was 7.6% and 33% of faces landed within five percent.
Golden ratio face questions
What is the golden ratio face test measuring?
+
It measures eight distances on your face and divides them into eight pairs. Face length against face width, mouth width against nose width, the three facial thirds against each other, the width of the face in eye widths, the gap between the eyes against one eye width, and nose width against that same gap. Each pair has a classical target, either 1.618, or 1.000 for the parts of the canon that ask for two equal measurements, or 5.000 for the rule of fifths. The test reports how far each pair sits from its target, with a sign, so a result of plus six percent means that ratio came out six percent larger than the classical value.
Is there such a thing as a perfect golden ratio face?
+
Not in any measured sense that we can back up. We ran this same test over 43 AI generated faces and not one of them landed within five percent of the target on all eight ratios. The best face in the set still averaged about ten percent away, and the widest was over twenty percent away. The classical values describe a drawing convention from art history, not a distribution that real faces cluster around, so treat a large gap as normal rather than as a fault.
Why does my upper third always look too short?
+
Because the face mesh has no hairline point. The topmost landmark sits on the upper forehead, below where most hairlines start, so the upper third is measured short and then padded by a flat twelve percent estimate. Across our 43 faces the upper third came out about forty three percent under the equal thirds target, every single time, which is the signature of a measurement problem rather than of forty three faces with unusual foreheads. The two ratios that depend on it are flagged in the results table.
Does a golden ratio face calculator say if someone is attractive?
+
Ours does not, and we would not trust one that did. What this page computes is the distance between a handful of your proportions and a set of numbers borrowed from classical art. Attractiveness is not a distance measurement. It varies between people, between cultures and between decades, and no landmark model has access to any of that. There is also a practical problem with the claim: since no face in our own set landed within five percent of the targets on all eight ratios, a tool that scored attractiveness this way would have to mark every real face down. The reason to look at the numbers is curiosity, or to understand what a stylist means when they talk about balancing the thirds of a face before a haircut.
What are the facial thirds and how are they measured here?
+
The face is split into three bands: hairline to brow line, brow line to the base of the nose, and the base of the nose to the bottom of the chin. The classical canon asks for all three to be about equal, so the target for each pair of thirds is 1.000. This page reports the lower third against the middle third, and the upper third against the middle third, separately, because they fail in different ways. The lower to middle pair is honest. The upper pair leans on the estimated hairline described above.
Is my photo uploaded when I run the test?
+
No. The face model downloads once from this site and then runs inside your browser tab. The measurement happens on pixels that are already in your device memory, nothing is written to a server, nothing is stored in cookies or local storage, and the page is not allowed to open a connection to anywhere except this site. The face model has a usage ping built into it that would call Google roughly once a minute, and that call is stubbed out in code and blocked by the site's content security policy as a backstop. Closing the tab discards everything. If you would rather not hand over a photo at all, the example button runs the whole test on one of our AI generated faces instead.
What 43 AI generated faces actually score
We ran this test over every face we hold: 22 AI generated test faces and the 21 AI portraits used elsewhere on this site. No real people, no scraped photos. The median face missed the classical values by 13.8% on average, the closest by 10.3% and the furthest by 21.8%. 0 of them landed within five percent on all eight ratios.
| Ratio | Median miss | Direction | Within 5% |
|---|---|---|---|
| Face length to face width | 15.9% | below target | 0% |
| Mouth width to nose width | 4.4% | below target | 56% |
| Chin to lips, against lips to nose | 7.4% | above target | 40% |
| Lower third to middle third | 6.2% | above target | 42% |
| Upper third to middle third | 43.5% | below target | 0% |
| Face width in eye widths | 2.3% | above target | 77% |
| Eye gap to eye width | 20.9% | above target | 0% |
| Nose width to eye gap | 7.6% | above target | 33% |
Measured 2026-09-16 with the detector this page ships, by running scripts/dataset-stats.mjs over the image sets in this repository.
Leonardo did not use the golden ratio on faces
Almost every page about facial phi opens with Leonardo. It is not true, and the correction comes from inside the field. Naini's 2024 review in Maxillofacial Plastic and Reconstructive Surgery states it plainly: "Contrary to popular speculation, Leonardo da Vinci did not employ the golden ratio in any of his human-proportional drawings." He knew the ratio well enough to draw the polyhedra for Pacioli's book about it, and never once applied it to a face.
You can check the primary source yourself. The text Leonardo transcribed around Vitruvian Man is Vitruvius, Book III: the face is "a tenth part of the whole height," the head "an eighth," the foot "one sixth," and the face divides into three equal parts from chin to nostrils, nostrils to brow, brow to hairline. Tenths, eighths, sixths, thirds. Whole numbers throughout. There is no 1.618 and no irrational number anywhere in it. Naini notes the same of the drawing: "the proportions of the figure do not actually match it, and Leonardo's own text above and below the image only mentions whole number ratios."
Nor did Pacioli claim what he is credited with. His 1498 De Divina Proportione is "predominantly on mathematics and perspective in art," and he did not argue the ratio was the secret of beauty. Euclid, further back still, "merely regards it as a rather general mathematical ratio" with "no relationship to beauty or any other aspect of the natural world."
Where the modern version actually came from
The phi mask that launched a thousand tools was built by a cosmetic surgeon who described his own method in a 2002 interview: "I began merely studying faces of professional models and movie stars. By virtue of the fact that they get paid for being attractive, they seemed to be a good group to study." That is the entire sampling frame. The interview contains no sample size, no sampling procedure and no statistics of any kind, and the byline identifies him as founder and CEO of the company selling the analysis.
When other people tested it, it did not hold. A 2008 review in Aesthetic Plastic Surgery concluded the mask "best describes the facial proportions of masculinized white women as seen in fashion models" and "does not appear to describe 'ideal' face shape even for white women." Measured against 40 photographs of professional Black models across twelve classical ratios, only one matched phi and the rest differed significantly. Among 286 Malaysian adults measured with calipers, 17.1% conformed, and no association was found between conforming and rated facial appearance.
Naini also names the methodological trick that keeps phi alive, and it is the reason this page reports eight signed deviations instead of one verdict: "Admirers and supporters of the golden ratio begin with the ratio in mind, then attempt to prove its existence in the attractive human face. The choice of landmarks is often arbitrary, as is head positioning in photographs, making the genuine test and subsequent confirmation or refutation of the theory appear unverifiable." Pick your landmarks after you know your answer and phi appears everywhere. Fix them in advance, publish the gaps with their signs, and it does not.
The fWHR range everyone quotes is not in the paper they cite
Facial width-to-height ratio turns up constantly alongside golden-ratio content, usually with a population range of "1.6 to 2.15" attributed to Carre and McCormick's 2008 paper in Proceedings of the Royal Society B. We read that paper in full: methods, results, all of Table 1 and all four figure captions. The strings "1.6" and "2.15" do not appear in it, and it states no population range for fWHR anywhere. The only range in the entire text is about correlation coefficients, which "ranged from 0.17 to 0.51."
The only fWHR values the paper contains are two means from 88 undergraduates: women 1.80, men 1.86. No minimum, no maximum, no spread. The circulating range appears to have been read off the x-axis of a scatter plot and then repeated from site to site without anyone opening the paper.
It is also a different kind of measurement from anything on this page. fWHR is descriptive, from evolutionary psychology, defined as cheekbone width over the distance from brow to upper lip, and it was studied against aggression rather than beauty. It has no target value. There is no good fWHR. Phi-based ratios are the opposite: a number a face is said to should approach. Putting them side by side, as many pages do, treats a descriptive variable and a prescriptive ideal as the same kind of thing.
The authors' own caveat undercuts photo measurement generally, and applies to us as much as to anyone: the relationship they found "was partly influenced by the posture of the head in the photographs," some subjects were smiling, and they note that angry expressions lower the brow and raise the upper lip, "a pattern of muscle activity that increases the facial width-to-height ratio." Measured from a photo, that ratio is partly a measure of expression and head tilt rather than of bone.
Which facial third is longest, and why the canon almost never holds
The equal-thirds rule is the oldest item on this page, inherited from Vitruvius through the Renaissance. Measured on real people, it essentially never holds.
| Study | Population | N | Equal thirds held in |
|---|---|---|---|
| Jayaratne 2012 | Southern Chinese, Hong Kong | 103 | 0% of both sexes |
| Karaca Saygili 2016 | Turkish, ages 20-35 | 200 | not one participant |
| Bozkir 2004 | Turkish, ages 18-25 | 500 | 1 subject (0.4% of men) |
| Al-Sebaei 2015 | Saudi Arabian, ages 20-24 | 168 | not valid for the majority |
Which third is longest has no single answer either, and that is the more interesting result. In the Hong Kong sample the lower third was longest in 49% of men, while the middle third was shortest in 52% of women. In the Turkish sample of 200 the pattern reverses outright: the middle third was longer than the lower third in 71% of women and 65% of men. In the Saudi sample the lower third was longer in about 90% of both sexes. So "the lower third is the longest," repeated everywhere without a qualifier, is a majority tendency in some populations and simply false in others.
The Turkish authors draw the conclusion we would: these proportions "cannot be a guide for planning surgical or non-surgical treatments," and "the neoclassical canons may fit a few subjects and they do not represent the average facial proportions."
One definitional warning, because it affects the two thirds ratios above. Several of these studies measure the upper third from the hairline to the nasion, while the classical canon and most clinical teaching use the hairline to the glabella. They are not the same distance and their numbers should not be pooled.
The two ratios you should not trust yet
The face mesh has no hairline landmark. Its topmost point sits on the upper forehead, so face length is measured short and then padded by a flat 12 percent. That padding is a guess, and it is the same guess for a high hairline and a low one.
You can see it in the numbers above. The upper third came out 43.5% under the equal thirds target on the median face, and not one face in 43 landed within five percent. A real population would scatter around the target. A measurement that misses in the same direction on every face is a measurement problem. We hit the same wall building the face shape guide: the only faces our classifier reads as diamond are ones with hair over the temples, which narrows the forehead the same way. Until there is a hairline landmark, read the two flagged rows as rough.
The anthropometry literature agrees, which we did not expect. A 2025 landmark-reliability study of 50 standardised frontal photographs found that of 37 facial landmarks, a handful "exhibited lower reliability" and named the trichion, the hairline point, among them. And in two independent published tables of facial thirds, one Turkish and one Saudi, the upper third carries by far the largest standard deviation of the three: roughly 7.5 to 9.6 mm against 4.1 to 5.2 mm for the middle third. The least reliable landmark sits at the top of the least stable measurement, in other people's data as well as ours.