This face symmetry test mirrors your own photo in the browser, then measures six feature groups separately instead of handing back a single symmetry score. When a difference is smaller than the measurement can resolve, it says so.
Front-facing photo, even light, hair off the forehead
Your two mirrored composites, left half onto itself and right half onto itself, replace this diagram once a photo is measured.
Six groups, seventeen point pairs, shown on the diagram above. Each group is compared on two axes: how far each side sits from the midline, and how far one side sits above the other. Here is what each group is actually looking at, and what tends to move it for reasons that have nothing to do with the face underneath.
Three pairs per brow: the outer end, the middle of the arch and the inner end. The inner ends are the interesting ones, because the muscle that lifts a brow works on one side at a time and most people have a habitual side.
Brow grooming defeats this row faster than anything else on the page. If the arches were shaped separately, the middle pair is measuring the threading, not the bone. A raised brow at the moment of the photo does the same thing.
Four pairs: both outer corners, both inner corners, and the centre of the upper and lower lid on each eye. Corner positions are the stable part. The lid centres move with how open the eye is, which is why they contribute the height difference rather than the sideways one.
Squinting into a window, or one eye closing slightly more than the other because of the flash, will move the lid pair by more than a real lid difference would.
Two pairs across the nostril wings, plus the nose tip measured against the midline separately. The wings are a good pair to compare because they sit on cartilage that does not move with expression.
Almost nothing in a photo moves them except a turned head, which is why the nose row is usually the smallest number in the table and the first one to fall under the measurable floor.
Two pairs: the mouth corners and the two peaks of the upper lip. The corners carry most of a mouth's asymmetry, and they are the pair people notice in their own photos.
Any expression at all ruins this row. A smile is almost never even, and even a slight one will make the mouth the largest number on the page. Take the photo with the mouth relaxed and closed.
Two pairs: the widest point of the face on each side, which sits at roughly eye level rather than on the apple of the cheek, and a point lower on each cheek. Together they describe how the middle of the face is balanced.
Hair falling over one temple shifts the outer point on that side, and side lighting flattens one cheek and shadows the other, which moves the lower point. Pull the hair back and light the face evenly.
Four pairs down each side of the jaw, from the corner below the ear to the point where the jaw turns toward the chin. This is the group the model is least certain about, because the jaw edge is where a face meets shadow.
A beard blurs the edge the model is looking for, and a head tilted up or down foreshortens both sides unevenly. Treat a large jaw number on a bearded or tilted photo as unproven rather than real.
Faces grow in two halves that never quite match, then spend decades chewing on one side, sleeping on one side and raising one brow more than the other. So the interesting question is not whether a face is asymmetrical. It is how much, and where.
Here is what we get on our own set of 43 AI generated faces, which describe the generator and this tool rather than people. Brows and cheekbones move the most, the nose the least, and the spread between the steadiest face and the least steady one is about 16 times. These are generated faces rather than a population sample, so read them as a sense of scale, not as norms.
| Feature | Median difference | Against the floor |
|---|---|---|
| brows | 0.68% | usually measurable |
| eyes | 0.44% | usually inside the noise |
| nose | 0.21% | usually inside the noise |
| mouth | 0.50% | usually inside the noise |
| cheeks | 0.72% | usually measurable |
| jaw | 0.64% | usually measurable |
Measured on 2026-09-16, by running this same tool over every AI generated face in our set and recording what it returned. Median measurement floor across those faces: 0.58% of face width.
Everything here is measured on a flat image. A cheek that projects further forward reads the same as one that does not.
Head angle is corrected from the model's pose matrix, but a lens still compresses the far side. Square up to the camera.
A smile is rarely even. The mouth row will overstate the difference on any photo where the mouth is not relaxed.
Hair over one temple moves the cheekbone point on that side, and a beard blurs the jaw points the model is least certain about.
Sudden one sided change in a face is a medical question, not a styling one. This page is not a screening tool.
Below about 0.58% of face width the method cannot separate a real difference from its own wobble.
Other symmetry tools combine features into one number using percentage weights: jaw 35%, eyes 25%, and so on. We were asked for a defensible weighting and went looking for one. There is no published research establishing percentage weights for facial regions in a symmetry score. None. What does exist is a systematic review of how much asymmetry observers can actually detect, feature by feature, in millimetres, and it ranks the features in close to the opposite order.
| Feature | Noticed at | Model | Observers |
|---|---|---|---|
| Eyelid position at rest | 2.0 mm | 2D manipulated, female | Laypeople |
| Mouth corner | 3.0 mm | 2D manipulated, male | Laypeople |
| Brow height | 3.5 mm | 2D manipulated, male | Laypeople |
| Nose tip deviation | 4.0 mm | 3D model, male | Lay + clinicians |
| Chin deviation | 6.0 mm | 3D model, male | Lay + clinicians |
Read that ladder against a "jaw 35%" scheme and the problem is immediate: it puts the heaviest weight on the region observers tolerate most. The chin needs three times as much deviation as an eyelid before anyone notices. A scheme that weights jaw heaviest and eyes lighter is not a simplification of this evidence, it is the reverse of it.
The deeper reason we publish no weighting is that thresholds are not weights and cannot be converted into them. A threshold says observers stop noticing below some number. It says nothing about relative severity above that number, and nothing about how two features combine. The review that produced the ladder lists composite asymmetry as unresolved.
Seven studies, mostly 2D manipulations of a single young model, in increments as crude as 5 mm. A six millimetre chin threshold really means somewhere between 4 and 6.
One study found observers detect smaller asymmetries the longer they look. A threshold is not a constant, which is awkward for anything published as a fixed weight.
A nose deviating left is rated more asymmetric than the same deviation right. For the chin it reverses: right-sided deviation drew 30% more demand for surgery.
Even expertise does not behave the way a weighting would need it to. In one study of 90 raters, orthodontists, surgeons and laypeople all identified asymmetry at the same point, and all rated nose asymmetry as worse than chin asymmetry of the same size. In another, clinicians noticed at 2 mm where laypeople needed 4 mm. Those two results disagree, and that disagreement is part of the honest answer.
So this page reports six feature groups separately, with a floor under each, and does not roll them into a score. The review's own opening line is the better summary than any number we could print: "The perfectly symmetrical face does not exist and, if it did, it would be considered disconcerting and uncomfortable."
Threshold table from Wang, Wessels, Hussain and Merten, "Discriminative Thresholds in Facial Asymmetry: A Review of the Literature," Aesthetic Surgery Journal 2017. Corroborating rater studies: Chu et al. 2011 (30 observers), Meyer-Marcotty et al. 2011 (90 raters), Naini et al. 2012 (185 observers), Alrbata et al. 2020 (120 raters).
Asymmetry is not only universal, it has a direction. Three hundred healthy Czech adults, scanned in 3D and analysed with geometric morphometrics, showed statistically significant directional asymmetry in every sex and every age group tested. The pattern is consistent: the midline traces a shallow "C", the glabella leans left, the chin leans furthest left of all, and the overall effect leaves the right half of the face slightly larger. Landmark placement error in that study stayed under 1 mm.
So a face that comes out asymmetric is behaving like every face in a 300-person healthy sample. The population-level lean even has a direction you can predict.
This is the strongest single fact on the page, because it lets you put folklore and perception on the same axis. In 670 consecutive orthodontic patients, those with a one-sided crossbite had a measurably thicker masseter on the unaffected side: 11.42 mm against 11.15 mm. The difference is real and statistically significant. It is 0.27 mm.
Now set that against the ladder above, where the most sensitive feature on the face needs 2 mm before anyone notices. The chewing-side effect is roughly an eighth of the smallest visible asymmetry. The authors who measured it say so themselves: the difference "does not lead to any clinically -or aesthetically- noticeable asymmetry." A genuine effect, an order of magnitude below visibility, is the shape of most of what gets blamed for facial asymmetry online.
One prospective study of 71 patients found that people with a side preference had a lower upper-eyelid margin on that side. That is the whole finding. It showed no difference in brow position, no difference in lower-lid position, and it is cross-sectional with self-reported sleep position, so it cannot establish direction. The widely circulated claim that "64% show more aging on their sleep side" traces to press coverage of a review about wrinkles, not to any study reporting that number, and we could not find a primary source for it.
Posture is worse served. The nearest systematic review concludes that current evidence "is not sound enough" even for the related question it studied, and no review exists on posture and facial asymmetry specifically. "Bad posture causes facial asymmetry" is unsupported at review level, which is not the same as disproven, but is a long way from established.
One paper is cited for this everywhere, a 2018 study in JAMA Dermatology. We read it in full. It enrolled 27 people and analysed 16, all women, mean age 53.7, with no control group and no blinding of participants or instructors. Two of its nineteen scales reached significance, several moved the wrong way, one of the two "blinded" raters is the first author, and the exercise instructor is a co-author who founded the commercial regimen being tested.
But the decisive objection is simpler than any of that. We went through its full outcome list, sixteen items from forehead lines to neck volume, and it contains no symmetry item, no left-versus-right comparison and no laterality measure at all. The one study cited to show facial exercise improves symmetry never measured symmetry. Where symmetry genuinely is an outcome, it is in rehabilitation after a diagnosed nerve palsy, with drugs on board in both arms, which says nothing about a healthy face.
Everything above is about faces that have always been slightly uneven. A face that changes is a different matter, and it belongs with a clinician rather than with a measuring tool.
The distinction clinicians draw is about the forehead. American Family Physician's review of Bell palsy describes it as suspected with "acute onset of unilateral facial weakness or paralysis involving the forehead in the absence of other neurologic abnormalities," and lists the findings that point elsewhere: "bilateral involvement, sparing of the forehead, abnormal extraocular movements, hearing loss, tinnitus, or vertigo," noting these "indicate an upper motor neuron lesion or a lesion involving more than just cranial nerve VII."
That passage is written for a physician examining a patient. It is a reason to see one, not a test to run on yourself in a mirror, and this page is not going to turn it into a checklist. Sudden one-sided facial weakness is an urgent medical question. Take it to a doctor, not to a face measurement tool.
Sources: Hoffmannova et al., PLOS ONE 2023 (N=300 Czech adults, 3D) · Antonarakis et al. 2025 (N=670 orthodontic patients, ultrasound) · Tran et al., Ophthalmic Plast Reconstr Surg 2022 (N=71) · Zhang et al., PLOS ONE 2022 (systematic review) · Alam et al., JAMA Dermatology 2018 (27 enrolled, 16 analysed) · AAFP, "Bell Palsy: Rapid Evidence Review," 2023.
It places 478 points on your face in your browser, undoes the head angle using the model's own pose matrix, and finds the midline from twenty points that run down the centre of the face. Then it takes six groups of mirrored point pairs, brows, eyes, nose, mouth, cheekbones and jaw, and for each group it measures two things: how much further one side sits from the midline than the other, and how much higher one side sits than the other. Both are reported as a share of your face width, so the numbers do not change when you move closer to the camera.
Almost certainly not, and almost certainly nobody's. Across the 43 AI generated faces we measured, the median face showed a difference of about 0.52 percent of face width, and the widest was 3.24 percent. What is more interesting is the other end: on 6 of those faces we could not measure any feature difference above our own measurement noise. That does not mean those faces are perfect. It means the difference, if there is one, is smaller than this method can see.
It is the oldest trick in this corner of the internet. The page takes the half of your photo on one side of the midline, mirrors it, and joins it to itself, then does the same with the other half. You end up with two faces built from one half each. They often look like two different relatives of yours. That is partly real asymmetry and partly the camera, because a lens sees one half slightly closer than the other unless you are perfectly square to it. Treat the composites as a way to see where a difference sits, not as a measurement.
Because it can tell the difference between a small number and a meaningless one. Each photo is measured five times with tiny shifts, and the points move a little between passes. On this set of faces the median wobble was about 0.42 percent of face width per point, and a difference between two points can wobble more than one point alone. Anything below the floor shown next to your result is inside that noise, so the page says so rather than printing a confident number you should not act on.
Yes, easily. A head turned a few degrees puts one side closer to the lens, which widens it in the image, and this page corrects the head angle but cannot recover pixels the lens compressed. A smile is rarely even, so the mouth row will overstate the difference on a smiling photo. Hair falling over one temple shifts the cheekbone point on that side. Ring lights and window light also flatten one side and shadow the other, which moves the jaw points the model is least sure about anyway.
This page does not score attractiveness and will not tell you that a smaller number is better. The research people usually cite on symmetry and attraction reports small average effects across groups, not a threshold that an individual face passes or fails, and it was not measured with a phone camera and a landmark model. Our own numbers make the point another way: the spread between the most even and the least even face in our set is wide, yet the differences involved are mostly under one percent of face width, which is a couple of millimetres on a real head. What the numbers here are good for is practical. Noticing that your brows sit at different heights before a haircut, or working out why one side of your face photographs better than the other.