Methodology
How we measure, and where the measurement stops
1. Finding the face
The photo is passed to Google’s MediaPipe Face Landmarker, which returns 478 points on the face — the jaw outline, the eye corners, the brow, the nose, the lips. It runs as WebAssembly inside your browser tab, on a model file served from this site’s own domain, so neither the photo nor a request about it reaches us or any third party.
If no face is found, the tool refuses to answer. This sounds obvious and is not: naming a shape for a photo with no face in it is the single worst failure mode here, because a wrong answer looks exactly like a right one. A test in the build feeds a blank image to the analyser and fails if anything but a refusal comes back.
Head pose is checked before anything is measured. A face turned away from the camera foreshortens on one side, which makes a symmetrical face read as asymmetric and narrows the far cheekbone. Beyond the yaw and roll limits the result is flagged rather than quietly reported.
2. The four measurements
Face shape is a set of proportions, so every input is a ratio — a width against another width — and none is an absolute size. That is why the answer does not change when you stand closer to the camera.
- Height to width
- Face height (upper forehead to chin) divided by the widest point. Separates a long face from a short one — the oval/oblong distinction lives almost entirely here.
- Forehead to face width
- Forehead width as a fraction of the widest point. A narrow forehead against wide cheekbones is what makes a diamond, and a wide one is half of what makes a heart.
- Cheekbone to jaw
- Cheekbone width divided by jaw width. The higher this runs, the more the face tapers toward the chin.
- Jaw angle
- The angle the jawline turns through at the corner, in degrees. A low angle is a squarer, more defined corner; a high one is a softer curve.
3. The reference profiles
Each of the six shapes is defined as a centre value and a tolerance for all four measurements. Your face is scored against every profile by weighted standardised distance — how many tolerances away you sit on each measurement, weighted by how much that measurement actually distinguishes shapes — and the nearest profile wins.
When the winner and the runner-up are close, the result is presented as borderline between the two rather than as a confident single answer. Real faces do not sort into six boxes, and a tool that pretends otherwise is choosing tidiness over accuracy.
| Shape | Height : width | Forehead : face | Cheek : jaw | Jaw angle |
|---|---|---|---|---|
| Oval | 1.24 | 0.88 | 1.26 | 137° |
| Round | 1.13 | 0.88 | 1.2 | 141° |
| Square | 1.16 | 0.9 | 1.2 | 130° |
| Heart | 1.24 | 0.92 | 1.37 | 134° |
| Oblong | 1.34 | 0.88 | 1.24 | 137° |
| Diamond | 1.24 | 0.84 | 1.31 | 134° |
Tolerances are ±0.1 on height:width, ±0.07 on forehead:face, ±0.1 on cheek:jaw and ±10° on jaw angle, identical across all six profiles — otherwise the shape with the loosest tolerance would win at every extreme value. Weights: 29% height:width, 29% cheek:jaw, 24% forehead:face, 18% jaw angle.
4. Eye shape, on two separate axes
The eye tool reports two things that are usually mashed into one word. The first is opening proportion: eye width divided by eye height. Below 2.6 reads as round, above 3.3 as elongated, and between them as almond.
The second is corner tilt — the angle from the inner corner to the outer one. This is the measurement that was wrong, and it is worth stating plainly because the correction is the best evidence on this page that the numbers here are checked against reality rather than against intuition.
Correction · August 2026
The neutral band was centred on 0°, a perfectly level eye. Human canthal tilt actually peaks near +4°, with the typical range running +2° to +7°. Centred on zero, the tool called 74.7% of people upturned and only 4.4% downturned — so almost everyone got the same answer, and the minority who genuinely have downturned eyes were handed advice for the opposite problem.
The direction of the error showed up on eight test faces here. It was confirmed against 27,343 real analyses on the sibling site running the same code before the band was moved — whose median tilt sits almost exactly on +4°. The band is now +2° to +7°, and the split reads 16.4% upturned, 58.3% level, ~25% downturned.
What this cannot do
Four limits, stated before you rely on the answer rather than after you complain about it.
It cannot tell you if your eyes are hooded or monolid
Both are defined by the eyelid crease, and the 478-point mesh has no landmark on the crease — there is nothing there to measure. Any tool reporting hooded or monolid from a single photo is inferring it, not measuring it. We say so on the eye page rather than guessing.
It cannot see your hairline
The mesh's topmost point sits on the upper forehead, roughly 15% of face height below where hair actually starts. Face height is measured through that visible window and the reference values above are calibrated to it, which is why a number here will not match one you take with a tape measure to your hairline.
It measures the photo, not you
A camera close to the face widens the nose and cheeks; a tilted head shortens one side. Pose is checked and flagged, but a straight-on photo in even light is still the difference between a reliable answer and a plausible one.
It is not a judgement, and there is no score
No shape is better than another shape. Nothing on this site returns a rating, a percentile or a comparison, because the question 'what suits me' has a useful answer and the question 'how do I rank' does not.
Where the styling advice comes from
The geometry is reproducible; the advice attached to it is craft knowledge. It reflects the settled consensus among working hairdressers and opticians — balance width against length, soften a strong angle rather than repeat it, keep volume off the widest point — and it is written per shape rather than templated.
For glasses, one rule outranks frame shape entirely and is printed above every frame list: the frame must fit the width of your head. A perfectly-shaped frame that pinches at the temples is the pair you stop wearing.
None of this is medical advice. If something on this site conflicts with what your stylist says while looking at your actual hair, your stylist wins — they can see the texture, density and growth pattern that no photo measurement reaches.
Method last reviewed 2026-08-19. Found something wrong? contact@faceshapereport.com — see who writes this.