The golden ratio face test
What it measures
A golden ratio face test lays a template over your photograph and gives you a percentage back. This page is about where that percentage comes from, and what happened when the claim behind it was checked against faces that had actually been measured.
Also searched as golden ratio face calculator, golden ratio face test, phi mask, Marquardt mask and facial golden ratio.
What the test is doing
Phi is 1.618. The face version of it comes from two traditions.
The neoclassical canons are drawing rules out of Renaissance workshops. The face divides into three equal parts; the eye is one fifth of the face width; the mouth is one and a half times the width of the nose. The Marquardt phi mask is a twentieth-century wireframe of golden-ratio rectangles and decagons, which a portrait gets scaled to fit.
A calculator takes your landmarks, measures how far they land from one of those templates, and turns the distance into a score.
The templates are real. The arithmetic is real. The step that does not hold is the one in the middle, where fitting the template is treated as evidence that a face is rated well.
What the canons look like when a population is measured
These are the five canons Canthion checks, with the share of each measured sample that actually conformed. Every figure is transcribed from the anthropometric literature and appears again on the measurement page it belongs to.
| The canon | Share of the sample that conformed |
|---|---|
| ICD = PFL (orbital canon) | |
| Alar width = ICD (orbitonasal canon) | |
| Mouth width = 1.5 × nose width (naso-oral canon) | |
| Equal thirds (33.3 / 33.3 / 33.3) | |
| Five equal fifths (20% each) |
A rule that four faces in ten follow, at home, is describing a minority.
The findings people cite
- Holland, 2008. Procrustes and thin-plate-spline analysis of the Marquardt mask found it ill-suited to non-European populations, and found that it best describes masculinised white women, which is the opposite of what raters prefer.
- Khoshab and colleagues, 2022. A PRISMA systematic review of 27 studies covering six ethnic categories found no ethnic or gender group that consistently approximates either the neoclassical canons or phi.
- Pallett, Link and Lee, 2010. Asked people to choose between faces whose spacing had been altered, and found a preferred layout that has nothing to do with 1.618.
Holland 2008, Aesthetic Plast Surg — doi.org/10.1007/s00266-007-9080-z · Khoshab et al. 2022, Aesthet Surg J — doi.org/10.1093/asj/sjab339 · Pallett, Link & Lee 2010, Vision Res — doi.org/10.1016/j.visres.2009.11.003
The layout people actually preferred
Pallett's experiment produced two numbers. Vertically, an eye-to-mouth distance of about 0.36 of face height. Horizontally, a pupil-to-pupil distance of about 0.46 of face width. Faces near those values got chosen most often. Neither number was derived from phi or from any canon; both fell out of the choosing.
Canthion scores both like any other proportion. Midface ratio carries the vertical, eye spacing against face width carries the horizontal, each against the reference group you pick and each with its own interval.
Why a phi score off one photograph is not a measurement
An uncalibrated photograph carries no scale, so only ratios survive it at all. Head pose is worse. A few degrees of yaw shortens every horizontal distance on one side of the face and lengthens it on the other, which moves every width ratio the mask gets scored on.
Then there are the landmarks, which are themselves estimates. MediaPipe Face Mesh has a published mean absolute error of 3.96% of inter-ocular distance per landmark (Kartynnik 2019). On a photograph with 300 px between the pupils, that is about 12 px of wander on every point.
So every value here is published with an interval. A conformity percentage quoted to one decimal place, with no interval and no stated tolerance, is reporting noise as precision.
What is measured here instead
The proportions the mask is assembled from, one page each, with the formula, the reference tables and the error bar:
478 landmarks and 24 proportions on your own photograph, in your browser. The overview is free and the photo is never uploaded. The full write-up of why the phi mask is refused here is on the golden ratio refusal page.
Common questions
Is there a golden ratio for the human face?
There is a number, 1.618, and a long tradition of drawing faces with it. What there is not is evidence that faces closer to it get rated better looking. Pallett, Link and Lee put it plainly in 2010: "there was never any proof that the golden ratio was special. As it turns out, it isn't." The 2022 review of 27 studies across six ethnic categories landed in the same place: no group consistently approximates either the neoclassical canons or phi.
What is a good score on a golden ratio face test?
It depends entirely on which template that site picked, how it scaled your photo to fit, and how it weighted each ratio. Two calculators will hand the same photograph different percentages, and neither publishes a margin of error. A number with no method and no interval behind it is not something to be above or below.
Does Canthion give a golden ratio score?
No. It measures the proportions the phi mask is built out of and grades each against the population you pick, with an error bar on every value. What it will not do is multiply them into one conformity percentage against a template that measured faces do not follow.
Where does the golden ratio face idea come from?
Two places. The neoclassical canons are drawing rules from Renaissance workshops: equal thirds, the eye as a fifth of the face width, the mouth one and a half times the nose. The Marquardt phi mask is a twentieth-century wireframe of golden-ratio rectangles and decagons that a portrait gets scaled to fit. Both were offered as ideals to draw toward. Neither was a description of what people look like.
Are the neoclassical canons useless, then?
Useful as reference points, useless as targets. Knowing that your lower third is larger than your middle third tells you something real about your face. It is also true of about nine in ten of the adults in the Saudi sample above. A canon describes a drawing convention. A norm describes a population. Only the second can tell you where you sit.
What should I measure instead?
The proportions themselves, each against a reference group, each carrying the uncertainty the photograph actually has. Facial thirds and fifths, midface ratio, face width to height, nose width, mouth width and eye spacing are all measured here from one frontal photograph, and every one comes back with the interval that landmark noise puts on it.
Related: facial symmetry test · face shape · attractiveness test · how bone-level measurements are taken from a photo · methodology · glossary · all 35 insight pages