How This Site Works
Method, rule set, thresholds, and what we haven't proven yet.
Not yet reviewed by a licensed cosmetologist — we’re looking for one. If you are one and spot something wrong, tell us at contact@knowyourfaceshape.com and we’ll fix it and credit you.
Face shape isn’t a score. There’s no better or worse shape — only styles that work with the one you have. We don’t sell glasses or hair products, and these picks aren’t influenced by any brand. Your photo never leaves your device. We can’t see it, and we never could.
Affiliate links
There are no affiliate links, sponsored posts or paid placements on this site. We carried exactly one affiliate link, between 15 and 17 September 2026, and took it back out. If a retailer link ever comes back, it will carry a disclosure next to it and be marked rel="sponsored" in the page source, so search engines read it as a paid link rather than an editorial recommendation.
The pipeline, in plain language
- 1 · Landmarks. MediaPipe Face Landmarker runs in your browser and returns 478 points on the face. We use ten of them: hairline, chin tip, both cheekbones, both forehead edges, both jaw corners and both chin sides.
- 2 · Units. MediaPipe normalises x by image width and y by image height — different units. We undo that first, otherwise the same face would measure differently in a landscape photo than a portrait one.
- 3 · Ratios. Five features: face length ÷ cheekbone width, forehead ÷ cheekbone, jaw ÷ cheekbone, the angle at the jaw corner, and chin width ÷ jaw width. These landmarks are not the points a tape measure touches, so the ratios get converted before they are compared with anything — see Two measurement scales below.
- 4 · Distance. Each feature is compared to each of the seven prototypes in units of its own spread, weighted, and summed into one distance per shape. The smallest distance wins.
- 5 · Honesty. The confidence you see is the raw output of that comparison, rounded. It is never clamped to look better, and when it is low the page says so in yellow instead of hiding it.
The full rule set
The same constants the code runs on — this page imports them, it does not retype them.
Oval
- Face length ratio (face length ÷ cheekbone width) is at least 1.26 and at most 1.64 → your face is about 1.5× longer than it is wide
- Forehead ratio (forehead width ÷ cheekbone width) is at least 0.81 and at most 0.99 → your forehead is slightly narrower than your cheekbones
- Jaw angle (angle at the jaw corner) is at least 114° and at most 130° → your jawline is rounded rather than angular
- Jaw ratio (jaw width ÷ cheekbone width) is less than Forehead ratio (forehead width ÷ cheekbone width) → your jaw is narrower than your forehead
- Chin taper (chin width ÷ jaw width) is at least 0.36 and at most 0.54 → your chin is not sharply pointed
Round
- Face length ratio (face length ÷ cheekbone width) is at least 0.86 and at most 1.24 → your face is about as long as it is wide
- Jaw angle (angle at the jaw corner) is at least 124° and at most 140° → your jawline is soft with no sharp corners
- Chin taper (chin width ÷ jaw width) is at least 0.46 and at most 0.64 → your chin is broad rather than pointed
Square
- Jaw angle (angle at the jaw corner) is at least 100° and at most 116° → your jaw corners are sharply angled
- Face length ratio (face length ÷ cheekbone width) is at least 0.86 and at most 1.24 → your face is about as long as it is wide
- Forehead ratio (forehead width ÷ cheekbone width) is at least 0.86 and at most 1.04 and Jaw ratio (jaw width ÷ cheekbone width) is at least 0.83 and at most 1.08 → your forehead, cheekbones and jaw are all close to the same width
- Chin taper (chin width ÷ jaw width) is at least 0.46 and at most 0.64 → your chin is broad rather than pointed
Oblong
- Face length ratio (face length ÷ cheekbone width) is at least 1.46 and at most 1.84 → your face is noticeably longer than it is wide
- Forehead ratio (forehead width ÷ cheekbone width) is at least 0.86 and at most 1.04 and Jaw ratio (jaw width ÷ cheekbone width) is at least 0.80 and at most 1.05 → your forehead, cheekbones and jaw are close in width
- Chin taper (chin width ÷ jaw width) is at least 0.41 and at most 0.59 → your chin is not pointed
Heart
- Forehead ratio (forehead width ÷ cheekbone width) is at least 0.96 and at most 1.14 → your forehead is the widest part of your face
- Jaw ratio (jaw width ÷ cheekbone width) is at least 0.59 and at most 0.85 → your jaw is clearly narrower than your cheekbones
- Chin taper (chin width ÷ jaw width) is at least 0.21 and at most 0.39 → your chin tapers to a point
Diamond
- Forehead ratio (forehead width ÷ cheekbone width) is at least 0.71 and at most 0.89 → your cheekbones are wider than your forehead
- Jaw ratio (jaw width ÷ cheekbone width) is at least 0.59 and at most 0.85 → your jaw is narrower than your cheekbones
- Chin taper (chin width ÷ jaw width) is at least 0.21 and at most 0.39 → your chin tapers to a point
- Face length ratio (face length ÷ cheekbone width) is at least 1.21 and at most 1.59 → your face is longer than it is wide
Triangle
- Jaw ratio (jaw width ÷ cheekbone width) is at least 0.96 and at most 1.21 → your jaw is the widest part of your face
- Forehead ratio (forehead width ÷ cheekbone width) is at least 0.71 and at most 0.89 → your forehead is the narrowest of the three widths
- Chin taper (chin width ÷ jaw width) is at least 0.41 and at most 0.59 → your chin is not pointed
Every threshold, in one table
| Feature | Oval | Round | Square | Oblong | Heart | Diamond | Triangle | σ | weight |
|---|---|---|---|---|---|---|---|---|---|
| Face length ratio | 1.45 | 1.05 | 1.05 | 1.65 | 1.35 | 1.40 | 1.30 | 0.18 | 1.6 |
| Forehead ratio | 0.90 | 0.88 | 0.95 | 0.95 | 1.05 | 0.80 | 0.80 | 0.08 | 1.0 |
| Jaw ratio | 0.80 | 0.85 | 0.95 | 0.92 | 0.72 | 0.72 | 1.08 | 0.12 | 1.1 |
| Jaw angle | 122° | 132° | 108° | 118° | 120° | 120° | 118° | 8° | 0.8 |
| Chin taper | 0.45 | 0.55 | 0.55 | 0.50 | 0.30 | 0.30 | 0.50 | 0.08 | 0.5 |
These values were fitted on 12 September 2026 against the labelled photo set described below, replacing the literature-derived placeholders this site shipped with. How each prototype was placed: we took the ordering and spacing of the published anthropometric prototypes and rescaled them as a block into the range our photographs actually produced. Oval, Oblong, Round and Square had five or more labelled examples each, so their positions are anchored by measured group means. Heart, Diamond and Triangle had one, three and zero — those three sit where the literature ordering puts them, not where measurement does, and we would rather say so than let the table imply seven equally grounded rows.
The set is a calibration aid, not a validation set, and every label on it carries a single annotator’s judgment with no second independent pass. Read plainly: on that set these five measurements do not cleanly separate the seven labels. That is why every result on this site carries a confidence band and, when the runner-up is close, a blend hint — instead of a single answer presented as settled.
Two derived settings you will also see on result pages: a blend is flagged when the runner-up reaches 60% of the top score, and confidence below 45% switches the result header to a warning, with 60% and above counted as a confident read.
How our reference set was built
The proportions behind our shape rules were tuned on 12 September 2026 against a local calibration set of 50 frontal photographs of real people, collected from free-license stock libraries and used only on our own machines for measurement — none of these photographs appear on this site. Every photograph was labeled by a single annotator, with no second independent pass, so the labels carry one person’s judgment rather than a consensus. This set is a calibration aid, not a validation set: we publish no accuracy percentage derived from it and make no claim that the thresholds were verified against independent ground truth. Face illustrations shown on this site are synthetic and labeled as such.
Two measurement scales
Every number we publish — the table above, the typical ranges on each shape guide, the figures, the explanations on your result page — is stated on the tape-measure scale: the ratios you would get by putting a tape to your own face. The photo detector does not measure that way. It works from facial landmarks, and the landmarks it uses for face length and jaw width are not the points a tape touches, so its raw ratios fall in a noticeably narrower band than a tape’s would.
So we calibrate the classifier on the detector’s own scale and connect the two with one affine map per feature. Because that map scales the prototype centre and the standard deviation by the same factor, the standardised distance the classifier actually computes comes out identical on both scales — the table above is an exact description of the classifier, not a loose translation of it.
One consequence we would rather state than bury. On the photopath, the ratios on your result page are an estimate of what a tape measure would have read, converted from landmark geometry — not the raw landmark quotients, and the page says so next to the table. On the tape-measurepath, your own ratios are shown back to you as you entered them.
Seven shapes, not eight
We list seven shapes, not eight. Some styling guides use “inverted triangle” as a synonym for Heart and others treat it as a separate category — the boundary between the two is not consistent enough to classify against, so we keep it inside Heart and say so rather than inventing an eighth prototype.
Open data
The rule set on this page, and every dated comparison table our articles publish, are also available as data — one public repository on GitHub, under the CC BY 4.0 license. Open data has the full inventory: what is in the repository, how to cite it, and what the data does not prove.
References
- Farkas, L. G. (1994). Anthropometry of the Head and Face, 2nd ed. Raven Press.
- Farkas, L. G., Katic, M. J., & Forrest, C. R. (2005). International anthropometric study of facial morphology in various ethnic groups/races. Journal of Craniofacial Surgery, 16(4).
- Lugaresi, C., et al. (2019). MediaPipe: A Framework for Building Perception Pipelines. arXiv:1906.08172.
- Kartynnik, Y., et al. (2019). Real-time Facial Surface Geometry from Monocular Video on Mobile GPUs. arXiv:1907.06724.
- Milady Standard Cosmetology, 14th ed. Cengage Learning.
Corrections
If you are a licensed cosmetologist, a craniofacial researcher, or just someone who spotted a wrong number, write to contact@knowyourfaceshape.com Corrections are credited on the page they fix.