Does everyone have an asymmetric face?

Yes, in the ordinary samples on this page. The most even Western face still reads 0.18% of face width and the most even East Asian face 0.25%; the medians are 3.13% and 3.23%. A reading of 0.00% is the mathematical limit of this formula, not something observed in either sample, so no face here is perfectly symmetric.

What counts as a large left–right difference?

There is no clinical cut-off. The 5.21% tolerance used here is a display and scoring tolerance derived from twice the median deviation in an earlier sample, and it sits near the 68th percentile of the Western sample and the 70th of the East Asian sample. About three faces in ten exceed it. Above the tolerance is not abnormal, and below it is not perfect symmetry.

Can a phone camera make my face look asymmetric?

Yes. A turned head compresses one side, camera height changes the 2D projection, a wide lens held close to the face enlarges the nearer side, and hard side light moves the visible edges of the eyes and mouth. The tool rejects extreme pose and warns inside its gates, but the remaining pose still moves the number. Face the lens, hold the phone at arm's length and use even light.

Does a symmetric face score higher on the attractiveness test?

The evidence is mixed. A 2006 review reported a positive association, while large 2024–2025 studies found no independent effect of symmetry once averageness and femininity were controlled. The AI attractiveness test gives its symmetry row a weight of 3 out of 100 and scores it against the same 5.21% tolerance. This page reports no total and does not treat symmetry as attractiveness.

Does the test use different references for East Asian and Western faces?

No arithmetic changes by region. The seven pairs, the midline, the normalization, the tolerance and the row weight are identical; only the middle-50% band a reading is compared with differs. A switch above the reading selects that band, and the detail view prints both. The tool does not infer ethnicity from a photo, and it publishes no sex-specific bands because the Western pool has no sex labels and both pools are convenience samples.

Is this the same engine as the AI attractiveness test?

Yes. Both pages use the same MediaPipe Face Landmarker detector, the same 478 landmarks and the same measureAttractiveness function. The head page folds seven rows into a 0–100 index; this page opens the symmetry row and shows each of the seven pairs. Same landmarks, different presentation.

Why is there no 0–100 score on this page?

A 0–100 scale needs a population norm, and both sets here are convenience samples. The tolerance sits near the 68th and 70th percentiles, so a lower-is-better scale would crowd most faces into one middling band; the head page already carries an index; and on a single 2D photo the pose error is larger than the differences a rank would separate. Lower also does not mean better: the evidence on symmetry and attractiveness is mixed.

Is asymmetry a medical problem?

No. Most facial asymmetry is ordinary and subclinical. Sudden asymmetry, facial droop, double vision, pain or swelling is a different matter and needs a clinician. This page measures 2D landmarks and is not a diagnosis.

On-device · Free · No sign-up

Face Symmetry Test: How Symmetric Is Your Face?

Measure left–right differences at seven landmark pairs in your browser. See the average, a plain position word and how it sits against the Western or East Asian ordinary band — not just a verdict.

This page measures how far seven pairs of facial landmarks sit from the face's own midline, as a percentage of face width. It shows the average difference and one position word against a single ordinary band — not a 0–100 score. Each of the seven pairs and the widest pair are in the detail view. Asymmetry is ordinary: the median is 3.13% in the Western sample and 3.23% in the East Asian sample, and even the most even faces read 0.18% and 0.25%. A switch above the reading selects the Western or East Asian middle-50% band; only that band and its typical value change, never the seven pair readings. The photo is processed in your browser and is never uploaded, and this is not a diagnosis.

How the reading is measuredSeven pairs, no combined score.

Your photo stays in your browser.

Runs on Google MediaPipe in your browser. Your photo never leaves your device.

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 seven pairsEach pair's left and right distance from your midline, printed as measured. No combined score.

What the Face Symmetry Test Measures

The test compares seven pairs of points, one on each side of the face. For every pair it measures how far the left point sits from the face's midline and how far the right point sits, both as a share of face width. The difference between those two distances is that pair's reading.

The midline is the face's own iris line: the tool rotates the whole face until the line through the two iris centers is horizontal, then takes the midpoint of that line. When an iris cannot be read, the outer-eye line is used instead and the reading is marked with lower reliability.

The denominator is cheekbone width, landmarks 234 and 454, so the reading does not depend on photo resolution or distance from the camera. A value of 0.0313 means 3.13% of face width.

Only horizontal distances enter. Vertical asymmetry — one eye sitting higher, one mouth corner sitting higher — is not measured by this formula at all.

There is no combined score, no percentile and no rank. The default view prints one average and a position word that says whether it sits below, inside or above the selected sample's middle 50%; the seven pairs and the widest pair are printed in the detail view.

It is not a diagnosis. The reading is a 2D landmark projection, not a clinical measurement.

Seven Landmark Pairs on One Relative Scale

Each pair uses two of the 478 MediaPipe Face Landmarker points. The order below is fixed; the page never sorts the rows by size, because sorting would look like a ranking.

The seven landmark pairs and the MediaPipe points they use
Pair	Left index	Right index	What it measures
Temples	21	251	Outermost point of the upper face at the temple.
Outer eye corners	33	263	Where the upper and lower eyelids meet on the outside.
Inner eye corners	133	362	Where the upper and lower eyelids meet on the inside.
Cheekbones	234	454	Widest point of the face; also the denominator, face width.
Nose wings	98	327	Outer edge of each nostril.
Mouth corners	61	291	Where the upper and lower lips meet at the corner.
Jaw sides	176	400	Side of the jaw contour.

One worked example

midX = 500 px, faceWidth = 300 pxleft cheekbone at 620 px → dL = 120 / 300 = 0.400right cheekbone at 380 px → dR = 120 / 300 = 0.400pairDelta = 0.000left cheekbone at 625 px and right at 372 px → dL = 0.417, dR = 0.427 → pairDelta = 0.010

A pair contributes only when both of its points are readable. With five or six pairs the mean is taken over those pairs and the card says so; below five pairs the whole row is unavailable and the page asks for another photo. Every scored face in both samples had all seven pairs.

The widest pair is marked on the card and never scored on its own. In the samples below its median is 6.97% of face width in the Western pool and 7.71% in the East Asian pool — roughly double the average — because it is a single maximum, not a summary.

A difference printed as 0.00% means the two sides project equally far from the midline in this photo. It does not mean the face is perfectly symmetric, and it is not a target.

Asymmetry Is Normal: What the Number Does and Does Not Mean

Neither sample contains a perfectly symmetric face. The smallest average difference is 0.18% of face width in the Western sample and 0.25% in the East Asian sample; the medians are 3.13% and 3.23%.

The 5.21% tolerance used elsewhere on this site is not a normal/abnormal line. 31.59% of the Western sample and 29.78% of the East Asian sample exceed it, which places it near the 68th and 70th percentiles. It is a display and scoring tolerance, and this page uses it only to describe where a reading sits.

A single pair can look dramatic while the average does not: 61.80% of Western faces and 66.10% of East Asian faces have a widest pair above the 5.21% tolerance, while only about three in ten have an average above it. The card prints both figures for that reason.

The number is a horizontal projection of a single photo. It is not a 3D shape measurement and not a measure of fluctuating asymmetry, which needs repeated measurements to separate developmental noise from pose.

Nor is it a verdict. This page never says that a lower reading looks better. The evidence is mixed: a 2006 review reported a positive association between symmetry and attractiveness, while large 2024–2025 studies found no independent effect after controlling for averageness and femininity, and a 1995 manipulation found that faces made more symmetric were rated less attractive.

The claim under test is that a symmetric face is more attractive. The ten studies below are ordered by first publication date; each row is marked as supporting, qualifying or challenging that claim.

Dated facial-symmetry studies and what they bear on
Date	Study	Sample or method	Finding	Bearing on the claim
1995-02-01	Ferrario et al., A three-dimensional evaluation of human facial asymmetryJ AnatPubMed	80 white adults (40 women, 40 men); 3D digitization	The average face was significantly asymmetric in shape but not in size; asymmetry was unevenly distributed across the face.	Qualifies the claim
1995-07-01	Swaddle and Cuthill, Asymmetry and human facial attractiveness: symmetry may not always be beautifulProc Biol SciDOI	Synthetic faces manipulated toward symmetry	Faces made more symmetric were rated less attractive, not more.	Challenges the claim
1997-01-01	Severt and Proffit, The prevalence of facial asymmetry in the dentofacial deformities populationInt J Adult Orthodon Orthognath SurgPubMed	1,460 orthodontic and orthognathic patients	34% had clinically visible facial asymmetry.	Qualifies the claim
2005-02-01	Katsumata et al., 3D-CT evaluation of facial asymmetryOral Surg Oral Med Oral Pathol Oral Radiol EndodDOI	16 controls plus patients; 3D-CT	Proposed a millimetre asymmetry index from hard and soft tissue; the method is not a 2D photograph.	Qualifies the claim
2006-01-01	Rhodes, The evolutionary psychology of facial beautyAnnu Rev PsycholDOI	Review and meta-analysis	Symmetry, averageness and sexual dimorphism were reported to correlate with attractiveness.	Supports the claim
2008-12-01	Ercan et al., Facial asymmetry in young healthy subjects evaluated by statistical shape analysisJ AnatDOI	321 healthy adults (150 men, 171 women); 2D photographs; 20 points	Healthy faces were systematically asymmetric, with a left-side dominance; some distances were more asymmetric in women.	Qualifies the claim
2014-10-01	Pound et al., Facial fluctuating asymmetry is not associated with childhood ill-health in a large British cohort studyProc Biol SciDOI	4,732 British adolescents; 3D scans	Fluctuating asymmetry was not associated with childhood illness.	Challenges the claim
2015-11-01	Thiesen et al., Facial asymmetry: a current reviewDental Press J OrthodDOI	Review	Asymmetry is common, often subclinical and multifactorial.	Qualifies the claim
2024-01-01	Kleisner et al., Distinctiveness and femininity, rather than symmetry and masculinity, affect facial attractiveness across the worldEvol Hum BehavDOI	1,550 faces, 10 populations, 72 landmarks; Bayesian models	Symmetry had no effect on attractiveness; distinctiveness was negative and femininity positive.	Challenges the claim
2025-02-14	Lee et al., Further evidence that averageness and femininity, rather than symmetry and masculinity, predict facial attractiveness judgmentsSci RepDOI	Male and female faces; attractiveness ratings	Replication: averageness and femininity predicted ratings; symmetry and masculinity did not.	Challenges the claim

What changed, what did not

Changed: the question moved from whether symmetry correlates with attractiveness to whether it explains anything once averageness and femininity are controlled. The 2024–2025 cross-cultural studies answer no; the 1995 manipulation even found that more symmetric faces were rated less attractive.

What did not change: faces are measurably asymmetric (Ferrario 1995, Ercan 2008), asymmetry is common (Severt 1997, Thiesen 2015), and fluctuating asymmetry was not a health signal in a cohort of 4,732 British adolescents (Pound 2014).

The 2006 review is positive, but it predates the large controlled studies. The honest word for this literature is mixed, not proven.

The studies in the table do not measure this page's quantity. Ferrario, Katsumata and Pound used 3D imaging; Ercan used 2D shape analysis; this page uses horizontal 2D distances. Their numbers cannot be compared with these readings directly.

Kleisner 2024 and Lee 2025 are correlational. They do not license the reverse claim that a symmetric person cannot be attractive, and Pound 2014 only rules out a fluctuating-asymmetry link to childhood illness — not the fact that faces are asymmetric.

One Measurement, Two Ordinary Bands

Two convenience samples of ordinary stock portraits were measured with the same detector. The Western pool has 1,301 scored photos; the East Asian pool has 534 after 85 photos that appear in both pools were removed from the East Asian side.

The two middle-50% bands almost coincide: 1.68%–6.32% for the Western sample and 1.85%–5.73% for the East Asian sample. The medians differ by 0.10 percentage points. The visible difference is in the upper tail: the Western p90 is 9.58% and maximum 22.98%, against 7.66% and 12.73% in the East Asian sample. That is a difference between two samples, not a claim that one population is more asymmetric.

Nothing in the arithmetic changes with region. The seven pairs, the midline, the normalization, the 5.21% tolerance and the 3/100 row weight are identical in both reference profiles; only the band a reading is compared with changes. The switch above the reading therefore selects which ordinary band the headline compares against, and the detail view prints both bands side by side.

The page does not infer ethnicity from a photo, and it publishes no sex-specific bands. The East Asian pool does carry sex labels and the women's middle 50% sits 0.5–0.7 percentage points above the men's, but the Western pool has no sex labels and both pools are convenience samples, so no sex conclusion is drawn.

Two ordinary samples and their middle 50% bands
Sample	n	Median	Middle 50%	p90	Max	Above 5.21%
Western ordinary portraits	1,301	3.13%	1.68%–6.32%	9.58%	22.98%	31.59%
East Asian ordinary portraits	534	3.23%	1.85%–5.73%	7.66%	12.73%	29.78%

Why Two Photos of the Same Face Can Differ

The reading describes the photo as much as the face. Eight factors move it.

Photo factors that move the reading
Factor	What happens	What the tool does
Head turn (yaw)	One side is compressed and the other widened, which inflates the left–right difference.	Rejected above 0.30 IPD; warned from 0.12.
Head roll	Rotates the horizontal line; uncorrected, it would inflate the horizontal difference.	The face is first rotated on the iris line; rejected above 32°; warned from 12°.
Camera height (pitch)	Changes the 2D projection and the apparent cheekbone width.	Rejected outside 0.10–0.88; warned at or below 0.22 or at or above 0.72.
Lens distance and wide angle	A close, wide lens enlarges the nearer side and adds perspective distortion.	Not modelled; hold the camera at arm's length, not against the face.
Light	Hard side light moves the visible edge of the eye corners and mouth corners.	Not modelled; use even light without a single hard shadow.
Expression	Mouth corners and the jaw move.	Use a neutral expression.
Glasses, hair and beards	Cover the eye corners, temples or jaw sides.	Take them off or move them aside; with fewer than five readable pairs there is no reading.
Makeup	Changes the visible position of the eye corners and mouth corners.	Keep the same makeup when comparing two photos.

What a 15° roll does — and does not do

The demo in the result card rotates the stored landmarks by up to ±15° around the face centroid and re-runs the same measurement. It is not a new photo and not a new detection.

The reading stays inside 0.05 percentage points across the range, because the engine removes head roll on the iris line before any distance is measured. At 12° or more the reliability tag changes to moderate roll even though the number is stable: the number did not move, the trust did.

A turned head (yaw) is not simulated. Yaw changes the projection, and a rotated landmark set cannot reproduce that; the tool warns or refuses instead.

A photo that gives the most stable reading

Face the camera straight on, with both ears roughly equally visible.
Hold the phone at arm's length, not close to the face.
Keep the camera at eye level.
Use even, front-facing light without a hard shadow on one side.
Take glasses off and move hair away from the temples and jaw.

If the reading differs between two photos, the difference is usually the photo, not a change in the face.

How This Reading Relates to the Attractiveness Index

The AI attractiveness test builds a 0–100 index from seven rows. Its symmetry row uses these same pair readings, averages them and scores the result against the 5.21% tolerance with a weight of 3 out of 100 in both the Western v9 and East Asian v10 profiles. The tolerance, the weight and the underlying arithmetic are identical in both profiles.

This page shows no row score and no index. The panel receives the readings and a reliability assessment; it does not receive the index, the row score or the profile name.

That separation is deliberate. The index is a calibration against reference sets, not an aesthetic verdict, and a single 2D photo supports a description of seven horizontal differences more honestly than it supports a rank.

The readings do not depend on the reference profile: the same photo gives the same average, the same seven pairs and the same widest pair under either profile. The index can still differ, because other rows do.

Open the AI attractiveness test and its seven-row formula.

What This Test Cannot Tell You

Seven things are outside this measurement.

Vertical asymmetry. One eye or one mouth corner sitting higher is not measured; the formula uses horizontal distance only.
3D shape. It is a 2D projection of a face, not a volume or a surface comparison.
Fluctuating asymmetry. A single photo cannot separate developmental noise from expression, pose and lens.
Health. Pound 2014 found no association between fluctuating asymmetry and childhood illness across 4,732 adolescents; this reading is not a health signal.
Attractiveness. The evidence is mixed, and this page makes no beauty claim in either direction.
Ethnicity or sex. Western and East Asian are two reference samples to compare with; the page does not infer where you are from, and it publishes no sex bands.
Change over time. Two photos taken on different days differ by pose, lens and light as much as by the face; the page cannot date a change.

It is also not a diagnosis. Sudden asymmetry, facial droop, double vision, pain or swelling needs a clinician.

Frequently Asked Questions

Does everyone have an asymmetric face?: Yes, in the ordinary samples on this page. The most even Western face still reads 0.18% of face width and the most even East Asian face 0.25%; the medians are 3.13% and 3.23%. A reading of 0.00% is the mathematical limit of this formula, not something observed in either sample, so no face here is perfectly symmetric.
What counts as a large left–right difference?: There is no clinical cut-off. The 5.21% tolerance used here is a display and scoring tolerance derived from twice the median deviation in an earlier sample, and it sits near the 68th percentile of the Western sample and the 70th of the East Asian sample. About three faces in ten exceed it. Above the tolerance is not abnormal, and below it is not perfect symmetry.
Can a phone camera make my face look asymmetric?: Yes. A turned head compresses one side, camera height changes the 2D projection, a wide lens held close to the face enlarges the nearer side, and hard side light moves the visible edges of the eyes and mouth. The tool rejects extreme pose and warns inside its gates, but the remaining pose still moves the number. Face the lens, hold the phone at arm's length and use even light.
Does a symmetric face score higher on the attractiveness test?: The evidence is mixed. A 2006 review reported a positive association, while large 2024–2025 studies found no independent effect of symmetry once averageness and femininity were controlled. The AI attractiveness test gives its symmetry row a weight of 3 out of 100 and scores it against the same 5.21% tolerance. This page reports no total and does not treat symmetry as attractiveness.
Does the test use different references for East Asian and Western faces?: No arithmetic changes by region. The seven pairs, the midline, the normalization, the tolerance and the row weight are identical; only the middle-50% band a reading is compared with differs. A switch above the reading selects that band, and the detail view prints both. The tool does not infer ethnicity from a photo, and it publishes no sex-specific bands because the Western pool has no sex labels and both pools are convenience samples.
Is this the same engine as the AI attractiveness test?: Yes. Both pages use the same MediaPipe Face Landmarker detector, the same 478 landmarks and the same measureAttractiveness function. The head page folds seven rows into a 0–100 index; this page opens the symmetry row and shows each of the seven pairs. Same landmarks, different presentation.
Why is there no 0–100 score on this page?: A 0–100 scale needs a population norm, and both sets here are convenience samples. The tolerance sits near the 68th and 70th percentiles, so a lower-is-better scale would crowd most faces into one middling band; the head page already carries an index; and on a single 2D photo the pose error is larger than the differences a rank would separate. Lower also does not mean better: the evidence on symmetry and attractiveness is mixed.
Is asymmetry a medical problem?: No. Most facial asymmetry is ordinary and subclinical. Sudden asymmetry, facial droop, double vision, pain or swelling is a different matter and needs a clinician. This page measures 2D landmarks and is not a diagnosis.

Method and Sources

Detector: MediaPipe Face Landmarker, 478 points, run locally in the browser. The symmetry row uses the fourteen points in the pair table above; the midline falls back from the iris line to the outer-eye line when an iris is unreadable.

Samples: ordinary stock portraits from Pexels. The Western pool processed 1,509 rows and produced 1,301 scored readings. The East Asian pool processed 793 rows and produced 619, of which 534 remain after removing 85 photos that also appear in the Western pool. A face is never counted in both bands.

Gates: both samples were collected with an earlier gate (yaw 0.20, roll 20°, pitch 0.15–0.80). The live tool uses a wider production gate (yaw 0.30, roll 32°, pitch 0.10–0.88), so these distributions describe the collected sets, not the live acceptance rate.

Statistics: population SD (÷ n); quantiles use linear interpolation at (n − 1) × p; readings are kept to four decimals in the tables. The 5.21% tolerance comes from 988 Western scored faces measured on 2026-10-20: twice the median symmetry deviation of 0.026071.

Sensitivity: removing the same 85 overlapping photos from the Western pool leaves n = 1,216 and moves every band endpoint by 0.08 percentage points or less, and every share by 0.3 points or less.

Script and snapshot: scripts/symmetry-stats.mjs recomputes these numbers from the local calibration files, which are not public; the checked-in snapshot is docs/ai-attractiveness-test/batches/wave-2-face-symmetry-test/symmetry-samples-2026-10-28.json. The tests read the snapshot, not the photos.

Average left–right difference by sample
Sample	n	Mean ± SD	p25	Median	p75	p90	Min	Max	Above 5.21%
Western, full	1,301	4.3568% ± 3.5339%	1.6764%	3.1273%	6.3185%	9.5833%	0.1834%	22.9791%	31.59%
East Asian, overlap removed	534	3.9375% ± 2.5787%	1.8481%	3.2282%	5.7343%	7.6644%	0.2514%	12.7308%	29.78%
Western, overlap removed	1,216	4.3909% ± 3.5898%	1.6808%	3.1266%	6.3573%	9.6561%	0.1834%	22.9791%	31.91%

Widest pair by sample — not scored
Sample	n	Mean	p25	Median	p75	p90	Above 5.21%
Western, full	1,301	10.3579%	3.6562%	6.9666%	14.0009%	23.2624%	61.80%
East Asian, overlap removed	534	9.4117%	4.0722%	7.7066%	13.1911%	18.4216%	66.10%

Five limits, restated

It is not a population norm. Both sets are convenience samples of stock portraits, so they describe these photos, not Western or East Asian people in general.
It is only a 2D horizontal projection. It is not 3D shape, not volume, and not vertical asymmetry.
It is not fluctuating asymmetry. One photo cannot separate developmental noise from expression and pose.
It is not a health signal. A cohort of 4,732 British adolescents showed no association between fluctuating asymmetry and childhood illness.
It is not an attractiveness verdict, and not a diagnosis. Sudden asymmetry, facial droop, pain or swelling needs a clinician.

The photo is processed in the browser and is never uploaded. A reading describes one photo, not a person.

Sources

Ferrario VF, Sforza C, Miani A, Serrao G. A three-dimensional evaluation of human facial asymmetry. J Anat. 1995;186(Pt 1):103–110. PMID 7649806. PubMed
Swaddle JP, Cuthill IC. Asymmetry and human facial attractiveness: symmetry may not always be beautiful. Proc Biol Sci. 1995;261(1360):111–116. PMID 7644543. DOI
Severt TR, Proffit WR. The prevalence of facial asymmetry in the dentofacial deformities population at the University of North Carolina. Int J Adult Orthodon Orthognath Surg. 1997;12(3):171–176. PMID 9511487. PubMed
Katsumata A, Fujishita M, Maeda M, Ariji Y, Ariji E, Langlais RP. 3D-CT evaluation of facial asymmetry. Oral Surg Oral Med Oral Pathol Oral Radiol Endod. 2005;99(2):212–220. PMID 15660095. DOI
Rhodes G. The evolutionary psychology of facial beauty. Annu Rev Psychol. 2006;57:199–226. PMID 16318594. DOI
Ercan I, Ozdemir ST, Etoz A, et al. Facial asymmetry in young healthy subjects evaluated by statistical shape analysis. J Anat. 2008;213(6):663–669. PMID 19094182. DOI
Pound N, Lawson DW, Toma AM, Richmond S, Zhurov AI, Penton-Voak IS. Facial fluctuating asymmetry is not associated with childhood ill-health in a large British cohort study. Proc Biol Sci. 2014;281(1792):20141639. PMID 25122232. DOI
Thiesen G, Gribel BF, Freitas MP. Facial asymmetry: a current review. Dental Press J Orthod. 2015;20(6):110–125. PMID 26691977. DOI
Kleisner K, Tureček P, Saribay SA, et al. Distinctiveness and femininity, rather than symmetry and masculinity, affect facial attractiveness across the world. Evol Hum Behav. 2024;45(1):82–90. DOI
Lee P, Li J, Rafiee Y, Jones BC, Shiramizu VKM. Further evidence that averageness and femininity, rather than symmetry and masculinity, predict facial attractiveness judgments. Sci Rep. 2025;15(1):5498. PMID 39953080. DOI