AI Face Analysis: Symmetry, Ratios, Shape, Age, Harmony

It also names the parts of that list your own photo could not support, before you scroll past them.

The Short Version

What Does a Face Analysis Measure?

Every number on this page starts as a map like this one: landmark points fitted to a single photograph and joined into a mesh. A distance between two points becomes a ratio, an angle, or a percentage. The scan fits 478 of these points inside your browser, and each metric reads from the same map, which is why the numbers move together rather than independently.

Nine rows below come back with a value and a unit. Five more are named without a value: some because a single front photo cannot support them, one because the number belongs to a different instrument entirely. Every row states which case applies, and links to the page that measures it in full where one exists.

The geometric measurements: symmetry, proportions and harmony

The geometric measurements come from distances between landmarks, expressed as a percentage, a ratio, or a difference from a fixed value. Symmetry compares paired points across a fitted midline. Thirds, fifths, facial width-to-height ratio and golden-ratio conformity each divide one distance by another. Harmony then measures how far those numbers spread from each other. Together they contribute six individual measurements to the manifest above.

The descriptive measurements: face shape, the eye region and the jawline

The descriptive measurements come back as a class or an angle rather than a ratio, and that distinction changes what confidence looks like. Face shape returns a class with a margin to the next-nearest class, so a face called oval by a narrow margin is a different claim than one called oval decisively. Canthal tilt, the eye region's angle, returns a sign and a roll correction rather than a class. A jawline reading works the same way, as its own angle at the corner of the jaw. This scan reports face shape and canthal tilt directly; a jawline reading runs on its own jawline rating page, built on the identical landmark set.

What a single front photo cannot measure

A single front photograph cannot support five things this analysis would otherwise report, each for a reason rooted in geometry rather than in effort. None of these improves with a better photo of the same kind.

• Projection and profile angles, which need a second view from the side.
• Three-dimensional volume, which a flat image cannot recover.
• Skin colour and texture, distorted by compression, lighting and makeup rather than by modesty.
• Anything about a face in motion, since one frame is one instant.
• Averageness and sexual dimorphism, which need a reference population rather than a better photograph.

A dedicated side-profile scan and a skin-quality guide are both planned additions to this list. Until either ships, both rows stay named rather than guessed at.

How the 478-Landmark Mesh Works

One face mesh does all the work behind every number on this page. The scan fits 478 points to your photograph inside your browser, and every component reads from that same map rather than from separate photographs. Each result then checks itself against its own tolerance before it renders, which is why some rows carry a value and others carry a named reason instead. The number shown here and the number its own dedicated tool returns are the same number, for the same photo and the same version. If they ever differ, that is a measurement error on our side, not a rounding choice.

1. Fit 478 landmarks to your photograph inside your browser, with no image ever leaving the device.
2. Read each metric from the points that measurement needs, using one mesh for all of them.
3. Check each result against its own published tolerance before it renders.
4. Render the manifest with every metric that passed, and name the reason beside every one that did not.

What each measurement needs from your photo

Different measurements survive different photographs. The table states, for each metric this scan can report, how far your head can turn before the number stops meaning anything.

The tightest tolerance on the page belongs to symmetry, at 2 degrees of yaw, matched to the strictest published limit found in this category rather than set past it. Face-shape class survives fifteen. That gap is the whole argument for gating each metric on its own terms instead of passing or failing an entire scan on one number.

Why One Scan Gives Nine Numbers and Not Nine Verdicts

These nine measurements come from one mesh, so they move together rather than independently. Turn the head two degrees and several numbers shift at once, because they share the same landmark points and the same fitted midline. A composite built from correlated numbers is a weighted sum, not a fact discovered in the data: the weights are a choice, and a different choice produces a different total from the identical photograph. Two tools that both print an overall score in the eighties are not measuring the same thing unless both publish the weights behind that total, and outside this site, none do. The research agrees with that structure. Symmetry, averageness and sexual dimorphism each carry a small share of rated attractiveness rather than one of them carrying most of it [2][8]. A single blended number would therefore hide the one fact worth keeping: the components disagree with each other more than a total lets on. This page prints no single face-analysis number, because a number like that would be our opinion of the weights, not a measurement of your face.

What Do I Look Like to Other People?

Raters agree with each other about facial attractiveness more than beauty is subjective implies, and less than a single number implies. Within one culture, agreement between raters runs high, and substantial agreement holds across cultures too [2]. That still leaves room for you: private taste accounts for close to half the variance in ratings of the identical face [3]. Agreement between cultural groups is real but consistently weaker than agreement within a group, and it leans on different cues depending on which group is rating [4]. Put together, what other people see is neither a fixed fact about your face nor pure guesswork on their part; it is a shared component and a private one, roughly split. A landmark measurement like the ones on this page describes geometry. It does not poll anyone's opinion, and it cannot report what a specific person thinks of you.

What the Evidence Says About Face Measurement and Attractiveness

Four claims bound what this page asserts, strongest evidence first, each stated as a claim, a named study, an effect size, and what the finding does not license.

Raters agree with each other more than beauty is subjective suggests. Langlois and colleagues, in a meta-analytic review in Psychological Bulletin in 2000, found within-culture agreement running around r = 0.90, with substantial agreement holding across cultures too [2]. That does not license one number as a fact about a face. Hönekopp, in the Journal of Experimental Psychology in 2006, found private taste accounts for close to half the variance in the same ratings [3]. Coetzee and colleagues, in PLoS ONE in 2014, found cross-cultural agreement real but weaker between groups than within them [4].

Symmetry, averageness and sexual dimorphism each predict rated attractiveness, and each predicts a small share of it. Rhodes, reviewing the field in the Annual Review of Psychology in 2006, reports all three as reliable but modest predictors [1]. Van Dongen, in Annals of Human Biology in 2011, found the symmetry association carries signatures of publication bias, meaning the true effect sits smaller than the published average [6]. Scott and colleagues, in PNAS in 2014, found dimorphism preferences shift by rater and by task rather than pointing in one direction [7]. Grammer and Thornhill measured the direct association in 1994 [5].

The golden-ratio mask claim is not supported. Holland, in Aesthetic Plastic Surgery in 2008, found the phi mask fits fashion models rather than the general population [9]. Pallett, Link and Lee, in Vision Research in 2010, found the ratios people rate as optimal sit near population averages, not at 1.618 [10]. Neither study licenses phi as a rule of beauty; both bound the claim to a narrow one.

The health-signal story is weaker than the category's marketing. Foo and colleagues, in Scientific Reports in 2017, found facial cues predict rated attractiveness far more reliably than they predict measured health [8].

Face Analysis, Facial Analysis and Face Scanners: Four Things That Share One Name

Four different things are called facial analysis, and this page performs exactly one of them: measuring facial geometry from a photograph.

Four industries kept the same two words, and none of them called first to ask. If a search engine sent you here for one of the other three, the row above is the correction, not an apology.

What This Analysis Does Not Infer: Personality, Ancestry, Health, Intelligence

This analysis measures geometry from a photograph, and geometry cannot support six things people sometimes search a face analyzer to learn. Each one is refused for a stated reason, not as a policy footnote.

• Personality or character: a face reading is not a measurement, whatever the video calls itself.
• Ancestry or ethnicity: the reason is bias in how these systems get built, not difficulty.
• Health or any medical status: a landmark mesh reads position, not tissue.
• Intelligence: no landmark correlates with it, and none is claimed to.
• Sexual orientation: not a geometric property of a face.
• Potential, including model potential: a projection this scan has no data to support.

What This Score Cannot Tell You

This returns measurements rather than a score, and measurements have their own limits. Six of them are named below.

The complete construction, every reference population and every failure mode live on how we measure, and where the measurement fails.

If thinking about your face has stopped feeling like curiosity, the National Alliance for Eating Disorders helpline is free and answered by clinicians.

The measurement's account of your face is complete. What follows is about the list of numbers you now have: which of them carry weight, and what to do with the rest.

You Have Every Metric Now: Which Ones Actually Moved Your Score?

A manifest is only useful once you know which rows carry weight and which are noise. The next three sections translate your numbers into the site's scale, tell you how to check a result before trusting it, and rank what actually moves these measurements against what only sounds like it should.

Where these numbers sit on the 1-10 scale

Each metric above describes where your photograph sits on that metric's own scale: a percentage for symmetry, a ratio for proportions, a spread for harmony. The site's 1-10 scale is a separate vocabulary that translates a full set of measurements into one interpretive band, and it is built and explained on its own page rather than repeated here. See what a 1-10 attractiveness score really means for the band definitions your numbers eventually feed. A metric printed here in its own unit and a band printed there in the 1-10 vocabulary answer two different questions about the same photograph.

How to check a scan before you believe it

1. Hold your phone's rear camera at arm's length, not the front camera close to your face.
2. Position the lens at eye height, facing you directly.
3. Light your face evenly from the front, and keep your expression neutral.
4. Compare the new unit values against the old ones, metric by metric, rather than comparing band labels.

A difference smaller than the tolerance published in the table above is noise, not a change; a difference larger than it is worth a second retake before you trust either result.

What actually changes them

Camera distance, lighting direction and head position move these numbers immediately, and they change nothing about your face while doing it [11]. Grooming, styling, sleep and hydration move some measurements modestly over days and weeks, mostly through what the mesh reads at the hairline, brow and jaw outline. The claims that do not survive contact with the tolerance table above, mewing, jaw exercises and bone-smashing among them, are named once here and not argued again. None of them move an adult skull, so none of them move a landmark this scan reads.

Is an AI Face Analyzer Accurate?

This analyzer is accurate in the reproducible sense: the same photograph and the same software version return the same measurement every time, which is a narrower claim than being right about your face.

Reproducibility is the claim this page defends, and the per-metric tolerance table above is part of that defense rather than an apology for it. A metric only renders once its own photograph conditions are met, so a number that does appear has already passed a named check. No validated accuracy percentage exists for this analyzer, or for any consumer face analyzer, because no study has measured one against a criterion strong enough to publish. Printing one anyway would be the easiest fabrication available on this page. The honest comparison is not a bigger number. It is a published unit, a published tolerance and a stated limit, on a page most face analyzers ship without any of the three. The full method, every reference population and the version history live on the methodology page linked above.

Free AI Face Analyzers: What to Ask of Any of Them

Five questions separate a measurement from a number with a shrug attached, and you can run all five against any face analyzer, including this one.

• The unit behind every number, not just the number itself.
• Whether your photo runs on your device or leaves it.
• Whether it refuses a photo it cannot measure, instead of guessing.
• Whether it names what it did not measure, row by row.
• Whether it is one payment or a recurring one.

Every free tool in the Attractiveness Report passes this checklist, which is a stronger claim than most of the category can make, and the reason it goes unstated above.

Thirteen numbered citations below are peer-reviewed and indexed in PubMed. Every effect size named above was read from the source paper, not from a summary.

Not medical advice. This is a measurement of a photo, not a diagnosis, a clinical assessment, or a verdict on you.