Face Shape Detector: Free, Private, Shows Its Measurements

Google's MediaPipe Face Landmarker finds 478 points on your face. From the outline points, the detector takes four measurements, each one tied to a trait that stylists and opticians use to define the shapes. Before measuring, it rotates the 3D landmarks back to a straight-on view and levels the eyes, so a slightly turned or tilted head doesn't read as a narrower face.

Face width is the distance across the widest outline points at cheekbone level. Every other width is expressed as a share of it, which is why your photo's size or distance from the camera doesn't change the ratios much, as long as you are far enough away to avoid close-up distortion.

None of the stylist or optician sources we read gives a numeric cut-off between shapes; the only figure in print is a hairstylist's rule of thumb that an oval face is about one and a half times as long as it is wide. So we built reference points ourselves. We measured Google's canonical face model, the average face MediaPipe is built around, and then moved each shape's reference in the direction its professional definition describes: wider forehead for heart, wider jaw for triangle, longer for oblong, sharper jaw corner for square, and so on.

Your four numbers are compared with all seven references at once. Each difference is put on a common scale, where 4 degrees of jaw angle counts the same as 0.03 of a width ratio or 0.06 of the length ratio, so no single measurement dominates. The closer your face is to a reference, the larger that shape's share. The shares add up to 100% across all seven, and the detector shows you the two largest.

Because the method is a comparison, not a yes-or-no test, it can't be fooled into false certainty by a face that sits on a boundary. If your face is halfway between the oval and oblong references, you will see two similar percentages, which is the honest answer.

Most wrong shape results come from the photo, not the face. The detector's quality check catches the worst problems and asks you to retake rather than guessing, but you will get a steadier answer if you set the photo up well in the first place. These steps come from our rulebook; each is labelled with where the convention comes from.

A result like Oval 52% and Oblong 39% is not a failure; it tells you your face sits near the line between those two shapes. The most common borderline pairs follow from the definitions: oval and oblong differ mainly in length, round and square differ mainly in the jaw corner, and diamond and heart both have a narrow jaw and differ in whether the forehead or the cheekbones are wider.

To see why the detector hesitated, open the result and switch on the Widths layer, then the Jaw layer. The shape card lists your forehead and jaw widths as a share of face width and your jaw angle, so you can see which measurement sits between the two references. If your length ratio is the deciding number, a hairstyle with volume on top or a fringe will push how your face reads one way or the other, so the borderline itself tells you where styling has the most effect.

The practical rule is simple. With a gap of 30 points or more between the first and second shape, follow the first shape's guide. With a smaller gap, read both guides and start with the advice they share. Stylists' rules for neighbouring shapes overlap a lot, so the shared advice is the safe bet.

Honestly: accurate enough to be useful and repeatable, not precise enough to settle a boundary case by itself. Three limits apply to every photo-based detector, including ours.

First, the landmarks have error. Google's model card for MediaPipe Face Mesh V2 reports an average landmark error of 2.06% to 3.85% of the distance between the eyes, depending on face region and mode, compared with 2.56% disagreement between trained human annotators marking the same faces. So the model is about as consistent as a person with a ruler, and no better.

Second, the photo moves the numbers. On our own test selfie, taken with the head turned about 10 degrees and smiling, the left and right halves of the face matched in width by 72% before our 3D pose correction and 95% after it. Without that correction, a turned head simply looks like a narrower face. To show how much is left after correction, the detector measures your photo plus three re-framed copies of it and reports the spread as a ± range.

Third, the shapes themselves are conventions with no agreed boundaries, so there is no ground truth to be accurate against at the edges. That is why we show percentages instead of a verdict, and why we never claim an accuracy figure that nobody has measured.

If a different app said heart and ours says diamond, there are usually three explanations. Most detectors report one label, so a face near a boundary gets whichever side of the line that photo happened to land on. Detectors don't draw what they measured, so a jaw line taken from hair or shadow goes unnoticed. And detectors use different landmarks and different thresholds, none of them published, so the same face can cross one app's line and not another's.

Two practical checks settle most disagreements. Run several photos through the same detector, rather than one photo through several detectors, and look at which shapes keep appearing in the top two. Then look at the Widths layer and ask the plain question the shapes are built on: which of my three widths is largest? That answer, more than any label, is what a stylist or optician will work from.

One more check is worth doing if the disagreement is between round and oval or oblong: look at your photo's quality notes. A slight smile, a chin tipped down or a camera held below the face all shorten or widen the lower face, and they are exactly the conditions our quality check flags. Retake with a neutral face and the camera at eye level before deciding which detector was right.

A shape is only useful if it changes what you ask for at the salon, the barber or the optician. These are the next steps we suggest, drawn from our sourced rulebook.