On-device · Free · No sign-up
Golden Ratio Face Test: How Close Are Your Proportions to φ?
Compare two facial ratios with the golden-ratio reference in your browser. See the numbers, the gap to φ, a proportion map and where they sit in a measured Western or East Asian reference sample.
Two ratios are measured from your photo: face height ÷ face width, and mouth width ÷ nose width. We show both, the gap to φ = 1.618, and where the first ratio sits in a measured Western or East Asian reference sample. A proportion map adds symmetry, jaw angle and model-defined three-band and five-column splits as readings. φ is a declared reference, not a law of beauty: in both samples, most faces sit below it.
Your photo stays in your browser.
Runs on Google MediaPipe in your browser. Your photo never leaves your device.
Choose the reference sample you want to compare with. This is a reference choice, not a determination of your ethnicity.
How it works
- 01Add a photoUpload a file or take one with your camera. The photo stays in your browser.
- 02We find 478 pointsMediaPipe maps the points on your face, on your device.
- 03Read two ratiosFace height ÷ face width, and mouth width ÷ nose width, each printed as measured.
What the Golden Ratio Face Test Measures
This page measures two ratios that are repeatedly connected to the golden ratio, prints them as measured, and compares only the first with a reference band built from ordinary portraits measured by the same detector.
The result card also draws a proportion map from those readings: your face frame, the φ frame and your jaw lines, with no photo on the image. Three extra rows sit beside it — symmetry, jaw angle and the model-defined three-band and five-column splits — and they are reported as readings only.
It does not produce a 0–100 score, a beauty rating or a rank against other people. The purpose is to show the readings and the size of the gap, not to turn φ into a verdict.
φ = 1.6180339887 is printed as a fixed comparison value. On this page it is a declared reference, not a law of beauty: it is there so you can see the distance, not so the distance can grade you.
A reading near φ is not better than a reading away from it. The band position below says only where a reading sits in the middle 50% of one measured sample.
The Two Ratios, the Extra Readings and the Landmarks Behind Them
MediaPipe Face Landmarker returns 478 points on a face. R1 and R2 read eight of them; the extra readings use additional points from the same model.
R1 = |y(152) − y(10)| ÷ |x(454) − x(234)|R2 = |x(291) − x(61)| ÷ |x(327) − x(98)|The landmarks are not the points a tape measure touches. R1 from the detector falls in a narrow raw band, so it is mapped to a 1.05–1.65 display scale before it is compared with a band. The map is affine: the quartile position does not change, while the standard deviation is stretched by about 3.8×. Every R1 band on this page is stated on that display scale.
R2 is not mapped. It is close to a tape-measure reading but slightly narrower: in the Western sample it sits 0.48 SD below the published anthropometric mean.
Three extra readings come from the same 478 points. Symmetry is the mean side difference across seven mirrored landmark pairs, divided by face width. Jaw angle is measured at each jaw corner between the cheekbone and the chin, and the left and right sides are reported separately. The three-band and five-column splits are model-defined, not the classical canon: the top band starts at landmark 10, which sits below the hairline, and the five columns are measured across the cheekbones (234 and 454), not at eye level. Their middle 50% is therefore not 33/33/33 or 20/20/20/20/20. All of these are printed as readings only.
Both ratios are 2D projections of a 3D face. Head pitch, camera distance, expression, makeup and hairstyle move the readings. That is a measurement limit, not a character judgement.
Nothing here measures skin, depth, age, sex or ethnicity. The detector does not infer them and the page does not ask for them.
| Landmark | MediaPipe index | Used for |
|---|---|---|
| Forehead top | 10 | Top of face height; the most stable point the model gives below the hairline, not a tape-measure trichion. |
| Chin tip | 152 | Bottom of face height. |
| Cheekbone left | 234 | Left side of face width. |
| Cheekbone right | 454 | Right side of face width. |
| Nose ala left | 98 | Left side of nose width. |
| Nose ala right | 327 | Right side of nose width. |
| Mouth corner left | 61 | Left side of mouth width. |
| Mouth corner right | 291 | Right side of mouth width. |
How the Western and East Asian Reference Bands Were Built
The bands describe two convenience samples of stock portraits, not populations. Each photo was measured with the same detector and the same landmark definitions.
Western sample: 1,509 rows were processed, 1,453 passed the quality gate, 152 were rejected for head turn, and 1,301 produced a score. The sample has no gender labels, so no gender split is used. Of these 1,301 photos, 85 also appear in the East Asian pool. Removing them leaves n = 1,216 and moves the R1 band endpoints by no more than 0.005; every conclusion below is unchanged. This page reports n = 1,301.
East Asian sample: 793 rows produced 619 scores; 16 had no face and 158 passed the quality gate but were rejected for pose. The same 85 overlapping photos were removed, leaving n = 534 — 325 women and 209 men. A face cannot belong to both reference sets.
R1 band = the p25–p75 interval of the sample. Tolerance = mean(p90 − p75, p25 − p10). A reading inside the band is the middle 50%; outside it, the distance to the nearest edge decays linearly over that tolerance. R2 has no band on this page. The extra rows use the same-sample p25–p75 only to locate the reading, not to score it.
| Sample | n | R1 mean ± SD | R1 p25 | R1 median | R1 p75 | R1 below φ | R2 mean ± SD | R2 below φ |
|---|---|---|---|---|---|---|---|---|
| Western reference | 1,301 | 1.3846 ± 0.2296 | 1.2370 | 1.3927 | 1.5370 | 85.6% | 1.5124 ± 0.1194 | 81.4% |
| East Asian reference | 534 | 1.3306 ± 0.1951 | 1.2092 | 1.3349 | 1.4567 | 92.9% | 1.5097 ± 0.1007 | 86.0% |
Gender split, East Asian sample only, for description and not for scoring: women (n = 325) R1 1.3463 ± 0.1880, median 1.3452; men (n = 209) R1 1.3060 ± 0.2032, median 1.3120. The women's median is 0.033 higher. The band is combined; this version does not use sex-specific bands.
External check, R1: published North American white adults give 1.3941 ± 0.0855; published Chinese adults give 1.2957 ± 0.0819. The Western 2D sample is 0.11 published SD from its external mean. The East Asian 2D sample is 0.43 published SD above its external mean, but the gap is not uniform: women are +0.80 SD and men are −0.05 SD. The direction also differs: in the published Chinese sample women have the lower R1, while in this 2D sample women have the higher R1. Equal 50/50 reweighting only moves the total gap from +0.43 to +0.37 SD. No correction is applied; the limitation is stated instead.
Sources: Farkas et al. 2005; Sarna et al. 2023
External check, R2: published North American white adults give 1.5802 ± 0.1409; published Chinese adults give 1.1661 ± 0.1092; Korean 3D samples give 1.090–1.132. The Western 2D sample is 0.48 published SD below its external mean. The East Asian 2D sample is 3.15 published SD above its external mean — an instrument difference, not a population finding; see the mouth-to-nose section below.
Sources: Farkas et al. 2005; Sarna et al. 2023; Kwon et al. 2021
The 2D samples are also wider than the published tape-measure samples: the R1 standard deviation is 2.7× the published Western SD and 2.4× the published East Asian SD. Pose, lens and expression widen a photo sample in a way a tape measure does not. Age is not controlled in either sample.
Phi Is a Declared Reference, Not a Law
φ = 1.6180339887 is a mathematical constant. Saying that a face matches it is a claim about preference, not a result that follows from measuring one face.
No study in the dated table below supplies a φ threshold. Population means sit away from φ. The pro-φ result comes from 14 selected actresses; the larger anthropometric and 3D studies do not reproduce it.
On this page φ is a declared reference, not a law of beauty. The gap is printed because it is easy to compute and easy to misread, not because a smaller gap is a better face.
There is no consensus that a face closer to φ is more attractive. The evidence table marks each study as supports, qualifies or challenges; the challenges are not a footnote to the supports.
What the Evidence Supports — and What It Does Not
The pro-φ result comes from 14 selected actresses; the larger anthropometric and 3D studies do not reproduce it. Rows are ordered by first publication date.
| Date | Study | Sample or method | Finding | Bearing on φ |
|---|---|---|---|---|
| 1985-03-01 | Farkas et al., Vertical and horizontal proportions of the face in young adult North American CaucasiansPlast Reconstr SurgDOI | 153 young North American white adults; 9 neoclassical proportions | Most proportions did not hold; vertical proportions performed worst. | Challenges φ |
| 1997-07-01 | Wang et al., Differences in horizontal, neoclassical facial canons in Chinese (Han) and North American Caucasian populationsAesthetic Plast SurgDOI | 206 Chinese Han and 103 North American white adults; horizontal proportions | Nose width equal to one quarter of face width in 51.5% of the Chinese group and 36.9% of the white group. | Qualifies φ |
| 2005-07-01 | Farkas et al., International anthropometric study of facial morphology in various ethnic groupsJ Craniofac SurgDOI | North American white adults; tr–me, zy–zy, ch–ch and al–al | Face height ÷ face width 1.394 ± 0.086; mouth width ÷ nose width 1.580 ± 0.141. | Qualifies φ |
| 2008-03-01 | Holland, Marquardt's Phi mask: pitfalls of relying on fashion models and the golden ratioAesthetic Plast SurgDOI | Review of the Marquardt φ mask and its fitting method | The method is flawed and the mask does not fit East Asian or sub-Saharan African faces; it closer resembles a masculinised European female face. | Challenges φ |
| 2009-08-01 | Mizumoto et al., Assessment of facial golden proportions among young Japanese womenAm J Orthod Dentofacial OrthopDOI | 74 Japanese women: 30 patients, 30 models and 14 actresses; 7 golden proportions | All 14 actresses met all 7 proportions; models and patients deviated on several. | Supports φ |
| 2009-11-06 | Pallett et al., New golden ratios for facial beautyVision ResDOI | Pairwise comparison experiments; eye-to-mouth vertical distance and interpupillary distance | The preferred positions were 36% of face length and 46% of face width, not φ. | Challenges φ |
| 2013-03-11 | Rossetti et al., The role of the golden proportion in the evaluation of facial estheticsAngle OrthodDOI | 400 young adults with 3D stereophotogrammetry; 60 rated; 10 ratios | Seven of ten ratios were significantly different from φ; the ratios were unrelated to attractiveness ratings. | Challenges φ |
| 2019-02-11 | Burusapat and Lekdaeng, What is the most beautiful facial proportion in the 21st century?Plast Reconstr Surg Glob OpenDOI | 16 Miss Universe Thailand and 16 Miss Universe contestants; 16 golden proportions | Golden proportions were not statistically supported in modern pageant faces. | Challenges φ |
| 2020-12-30 | Kwon et al., Three-dimensional photogrammetric study on age-related facial characteristics in Korean femalesAnn DermatolDOI | 192 Korean women; 3D photogrammetry by age group | Mouth width ÷ nose width ranged from 1.090 to 1.132, well away from φ. | Challenges φ |
| 2021-02-22 | Hwang and Park, The divine proportion: origins and usage in plastic surgeryPlast Reconstr Surg Glob OpenDOI | Review of the history and clinical use of φ | No evidence that architecture or painting used φ before Pacioli in 1509; the Marquardt mask does not match the north-west European average face. | Challenges φ |
| 2021-06-18 | Sarna et al., Craniofacial anthropometric norms in a selected Kenyan and Chinese populationPlast Surg (Oakv)DOI | 90 Chinese and 90 Kenyan adults; direct measurement | Chinese face height ÷ face width 1.296 ± 0.082; mouth width ÷ nose width 1.166 ± 0.109. | Qualifies φ |
| 2022-06-21 | Londono et al., Assessment of the golden proportion in natural facial esthetics: a systematic reviewJ Prosthet DentDOI | Systematic review of vertical facial proportions against the golden proportion | No significant association between the golden proportion and facial evaluation scores; no threshold was found. | Challenges φ |
What changed, what did not
Changed: measurement moved from photographs and tape measures to 3D imaging; samples extended from north-west European and North American groups to Chinese, Korean, Japanese, Malaysian, Turkish and Thai groups; and the conclusion moved from treating φ as the standard to treating it as one reference among many.
What did not change: no study gives a φ threshold; every population mean in the table sits some distance from φ; the supportive results come from small, selected actor or pageant samples; and measurement definitions differ enough that the values cannot be compared across studies.
Why Most Faces Sit Away From Phi
φ sits above p75 in all four region-by-ratio combinations on this page: Western R1 p75 1.5370, East Asian R1 p75 1.4567, Western R2 p75 1.5910 and East Asian R2 p75 1.5750, against φ = 1.6180339887.
That is why most faces sit below φ. In the Western sample 85.6% of R1 readings and 81.4% of R2 readings are below φ; in the East Asian sample the shares are 92.9% and 86.0%.
A reading below φ is not an incomplete face and not a failed test. It means φ is not the middle of these samples; it is a reference point placed above their middle 50%.
Why the Mouth-to-Nose Ratio Is Reported but Not Scored
On 1,057 scored East Asian portraits from the wider East Asian calibration set, nose width ÷ outer-canthal distance was 0.3759 ± 0.0284, 11–14% lower than the published Chinese range (0.425–0.438) and 16–21% lower than the published Korean range (0.446–0.476). Mouth width ÷ outer-canthal distance was about 0.568, 12–14% higher than the published Chinese range (0.498–0.509) and about 5% higher than the Korean value (about 0.539).
Sources: Sarna et al. 2023; Kwon et al. 2021
Both ends move at once: the detector reads the nose too narrow and the mouth too wide. That is an instrument difference, not a finding that East Asian faces have a different mouth-to-nose ratio. A correction factor would hide the bias and would also absorb smile and expression differences as if they were population differences.
No validated correction exists, so R2 is reported without a band and without a score. It is still shown with its gap to φ; the gap is a reading, not a grade.
In the Western sample R2 is also reading only: it can be read, but the 0.48 SD difference from published values is enough that treating it as a score would overstate what the detector knows.
What This Test Does Not Tell You
Four limits belong with every result on this page, not in a footnote.
- It is not a population norm. The reference samples are convenience samples of stock portraits, and the page does not claim that Western or East Asian people usually look a certain way.
- It is not a beauty verdict. Being inside the band means only that a reading is inside the middle 50% of one sample.
- It is not a diagnosis. The landmarks are 2D projections and move with pose, lens, expression and makeup.
- It does not guess ethnicity. You choose the reference sample; the page does not infer where you are from, and it does not store the choice with your photo.
The test also does not measure skin, depth, age, sex or ethnicity, does not upload the photo, and does not produce a combined score.
How This Test Relates to the Attractiveness Index
The head-page test combines seven proportion rows into one index. This page is narrower: it reports two ratios plus three extra readings and does not feed them into that index. The head page's golden-ratio row carries a weight of 20 in the Western v9 profile and 0 in the East Asian v10 profile; the readings here are not that row's subscore and cannot be substituted for it.
No combined score appears here. A face exactly at φ would fall outside both regional bands, so a single number would not mean what its name suggests.
Read the full seven-row formula on the AI attractiveness test.
Frequently Asked Questions
- What is a golden ratio face?
- It is a face described with proportions that are compared with φ = 1.6180339887. This page uses two readings: face height ÷ face width, and mouth width ÷ nose width. φ is a declared reference, not a law of beauty, and the page prints the readings instead of turning them into a grade.
- Is a face closer to phi more attractive?
- There is no consensus. The pro-φ result in the dated table comes from 14 selected actresses; larger anthropometric and 3D studies do not reproduce it, and several find ratios significantly different from φ. No study supplies a φ threshold.
- Why are there only two golden-ratio readings?
- R1 and R2 are the two ratios with definitions that stay stable under this detector. The extra rows — symmetry, jaw angle and the model-defined three-band and five-column splits — are printed as readings only, not as more golden ratios.
- Why doesn't this page give a single score?
- A face exactly at φ would fall outside both regional R1 bands, because φ sits above each band's p75. A regional-match score would therefore read low for the face the page is named after, and a golden-ratio score would not mean what its name suggests. The page shows both ratios, the extra readings and the band position instead.
- Which reference sample should I choose?
- Choose the sample you want to compare with. It is a reference choice, not a determination of your ethnicity, and the page does not infer where you are from.
- Does the test guess my ethnicity?
- No. There is no ethnicity classifier and the page does not ask for one. The region is a reference sample you select; the photo and the choice stay in your browser.
- Why is my mouth-to-nose ratio not scored?
- On 1,057 scored East Asian portraits, this detector's nose width reads narrow and its mouth width reads wide relative to published anthropometric means, leaving the ratio 3.15 SD above the published value. A correction factor would hide that instrument bias and absorb expression differences. The ratio is therefore reported as a reading only.
- Is this the same as the AI attractiveness test?
- No. The AI attractiveness test combines seven proportion rows into one index. This page reports two ratios plus three extra readings and does not feed them into that index. The head page's golden-ratio row has a different definition and is not the same number.
- Why do most faces sit away from phi?
- Because φ sits above p75 in all four region-by-ratio combinations on this page. In the Western sample 85.6% of R1 readings and 81.4% of R2 readings are below φ; in the East Asian sample the shares are 92.9% and 86.0%. That makes φ an upper reference point, not the middle of the samples, and a reading below it is not a failure.
Method and Sources
Detector: MediaPipe Face Landmarker, 478 points, run locally in the browser. R1 and R2 use eight points: 10, 152, 234, 454, 98, 327, 61 and 291. The symmetry, jaw and split readings use additional points from the same 478-point model. No image is uploaded and no server-side model runs.
Quality gate: the photo must contain a detectable face, the face must be large enough and inside the frame, and head pose must stay within the gate. A rejected photo produces a message instead of a reading.
R1 is mapped from the detector's raw ratio to a 1.05–1.65 display scale with an affine transform; R2 is reported raw. The band is the p25–p75 interval of the stated sample; tolerance = mean(p90 − p75, p25 − p10). The East Asian sample removes the 85 photos shared with the Western sample. No correction is applied to R2.
Statistics: standard deviations are population SD (÷ n); p10, p25, p75 and p90 use linear interpolation at (n − 1) × p; table values are rounded to four decimals. Dates in the evidence table are first publication dates from Europe PMC; the source list keeps the journal volume year.
The four limits above are part of the method, not fine print.
Sources
- Farkas LG, Hreczko TA, Kolar JC, Munro IR. Vertical and horizontal proportions of the face in young adult North American Caucasians: revision of neoclassical canons. Plast Reconstr Surg. 1985;75(3):328–338. DOI
- Wang D, Qian G, Zhang M, Farkas LG. Differences in horizontal, neoclassical facial canons in Chinese (Han) and North American Caucasian populations. Aesthetic Plast Surg. 1997;21(4):265–269. DOI
- Farkas LG, Katic MJ, Forrest CR, et al. International anthropometric study of facial morphology in various ethnic groups/races. J Craniofac Surg. 2005;16(4):615–646. DOI
- Holland E. Marquardt's Phi mask: pitfalls of relying on fashion models and the golden ratio to describe a beautiful face. Aesthetic Plast Surg. 2008;32(2):200–208. DOI
- Mizumoto Y, Deguchi T, Fong KW. Assessment of facial golden proportions among young Japanese women. Am J Orthod Dentofacial Orthop. 2009;136(2):168–174. DOI
- Pallett PM, Link S, Lee K. New "golden" ratios for facial beauty. Vision Res. 2010;50(2):149–154. DOI
- Jayaratne YS, Deutsch CK, McGrath CP, Zwahlen RA. Are neoclassical canons valid for southern Chinese faces? PLoS One. 2012;7(12):e52593. DOI
- Rossetti A, De Menezes M, Rosati R, Ferrario VF, Sforza C. The role of the golden proportion in the evaluation of facial esthetics. Angle Orthod. 2013;83(5):801–808. DOI
- Alam MK, Mohd Noor NF, Basri R, Yew TF, Wen TH. Multiracial facial golden ratio and evaluation of facial appearance. PLoS One. 2015;10(11):e0142914. DOI
- Burusapat C, Lekdaeng P. What is the most beautiful facial proportion in the 21st century? Comparative study among Miss Universe, Miss Universe Thailand, neoclassical canons, and facial golden ratios. Plast Reconstr Surg Glob Open. 2019;7(2):e2044. DOI
- Kaya KS, Türk B, Cankaya M, Seyhun N, Coşkun BU. Assessment of facial analysis measurements by golden proportion. Braz J Otorhinolaryngol. 2019;85(4):494–501. DOI
- Kwon SH, Choi JW, Kim HJ, et al. Three-dimensional photogrammetric study on age-related facial characteristics in Korean females. Ann Dermatol. 2021;33(1):52–60. DOI
- Hwang K, Park CY. The divine proportion: origins and usage in plastic surgery. Plast Reconstr Surg Glob Open. 2021;9(2):e3419. DOI
- Sarna K, Sonigra KJ, Ngeow WC. A cross-sectional study to determine and compare the craniofacial anthropometric norms in a selected Kenyan and Chinese population. Plast Surg (Oakv). 2023;31(1):84–90. First published 2021-06-18. DOI
- Londono J, Ghasmi S, Lawand G, Mirzaei F, Akbari F, Dashti M. Assessment of the golden proportion in natural facial esthetics: a systematic review. J Prosthet Dent. 2024;131(5):804–810. First published 2022-06-21. DOI