Eye shape detector, without the made-up buckets

This eye shape detector measures one thing a front photo can actually support: the height of your eye opening divided by its width, after correcting for head angle. It then tells you where that number sits among the faces we measured. What it will not do is hand you a label like narrow, almond or round, because we checked whether those labels survive contact with data and they do not. Most of the names people use for eye shapes describe the eyelid crease rather than the opening, and the crease is exactly what a straight-on photograph cannot resolve. So you get a measurement and a position, and an honest account of what is missing.

Because we measured the distribution and there is nothing there to cut. Across 43 faces the eye opening ratio runs from 0.22 to 0.41 in an unbroken continuum: the largest space between any two adjacent faces is 0.02, which is smaller than the measurement error on a single reading. That error is the decisive part. The per-face uncertainty is about 0.03, while the entire spread between faces is only 0.05. When we applied the three-bucket cut points this site used to ship, 67 percent of faces changed their label under one standard error. A label that flips for two faces in three is not describing the face.

The commonly listed types are almond, round, monolid, hooded, upturned, downturned, deep-set, protruding, close-set and wide-set. They are not one kind of thing, which is the part nobody says. Upturned and downturned describe the angle between the eye corners. Close-set and wide-set describe spacing between the eyes. Almond and round loosely describe the opening. Monolid, hooded and deep-set describe the eyelid and the bone around it. Those last three are the ones people most want identified and the ones a frontal photograph is least able to establish, because they are defined by a crease and a depth rather than by an outline.

Not reliably from the front, and this tool refuses to guess. A hooded eye is defined by a brow-side fold of skin that descends over the crease; a monolid is defined by the crease being absent or not visible. Both are statements about a fold, and a fold is a depth feature. From straight on, a shadow, a low camera angle, eye makeup along the lash line, or simply looking slightly downward all reproduce the appearance of a fold that is not there, and a bright even light can erase one that is. Tools that report hooded or monolid from a single frontal photo are reading the lighting as much as the lid.

MediaPipe's Face Landmarker places 478 points on your face inside your browser. This page reads eight of them: the two corners of each eye, and the centre of the upper and lower lid on each eye. Head rotation is undone first using the model's own pose matrix. Width is the corner-to-corner distance, height is the lid-to-lid distance, and the ratio is the average of the two eyes. It is averaged deliberately: the median difference between a person's two eyes in our set was 0.007, about a fifth of the measurement error, so reporting them separately would be reporting noise with a side attached.

Neither, and this page will not rank you. A percentile says where a number sits among the faces we measured; it does not say that one end of the distribution is preferable to the other. It is also worth knowing what the population is before reading anything into your position in it. Our reference set is AI-generated images, not photographs of people, so the distribution describes what an image model produces and how precisely this tool can measure it. A percentile against that is useful for understanding your own measurement. It is not a statement about how you compare to human beings.