How Face Shape Detection Works

Four readings describe any face, and the detector takes all four from the landmark points. Forehead width spans the widest part of the brow. Cheekbone width runs between the two most prominent cheek points, usually the widest line on the face. Jaw width measures across the mandible angle below the ears. And face length runs from the center of the hairline to the base of the chin.

The shape is a pattern in those four, not any single one. The detector divides face length by the widest width to get a ratio near 1.5 for long shapes and near 1.1 for compact ones, then compares the three widths to see which is broadest. Those two comparisons place almost any face before a final check settles the close calls. The measurement guide walks the same readings by hand.

The 468 points come from a machine learning model, not from rules a person wrote. Google MediaPipe Face Mesh was trained on a large set of annotated faces to predict where each landmark sits on a new photo, and it returns the full mesh in a fraction of a second. The model infers a 3D position for every point, which lets it place landmarks correctly even when a face is turned slightly.

That density is what makes the measurements reliable. With 468 points, the detector does not depend on any single landmark being perfect: it reads the widest cheek point, the mandible angle and the hairline center from clusters of nearby points, so a small error in one place barely moves the result. The model handles the hard part, finding the face and its features, and the arithmetic that follows is straightforward.

Turning the mesh into a label is plain geometry. The detector takes the four measurements, computes the length-to-width ratio, and checks which of the three widths is broadest. A high ratio with a soft, cornerless jaw reads oval; a low ratio with a hard corner reads square; a forehead-widest pattern reads heart. The remaining outlines follow the same logic, each fixed by its own pairing of ratio and widest point.

Most faces sit close to one shape, and the detector names it. Some fall between two, a 1.3 ratio between round and oval, for instance, and there the confidence score reflects the tie rather than forcing a false certainty. The classification mirrors the tables on the shape pages exactly, so you can check the reasoning yourself: run the numbers the detector shows against the face shapes hub and the same logic applies.

The entire pipeline executes on your device, which is a privacy decision as much as a technical one. MediaPipe ships as a WebAssembly module and a model file that your browser downloads once, then runs locally. There is no upload endpoint on this site, so the image is never transmitted, received or stored anywhere, and closing the tab discards everything.

That design has a practical upside too: results appear in seconds because nothing waits on a network round trip after the one-time model download. It also means the tool works the same on any modern browser without an account, a login or a server queue. The AI model page covers what gets downloaded and how large it is, and the privacy claim is load-bearing across the site.

The confidence figure is not a measure of photo quality or of how attractive a face is. It reports how cleanly your proportions match one shape versus its neighbors. A high score means the ratio and the widths point firmly at a single outline. A lower score means your face sits near a boundary between two shapes, which is common and not a fault.

Reading it that way makes the result more useful. A confident oval reading needs no second guessing, while a split reading between, say, heart and diamond is a signal to read both shape pages and take what matches. The score is honest about uncertainty by design, since a face genuinely between two outlines should not be labeled with false precision. The accuracy page explains what moves the score up and down.

Almost every weak reading comes from the photo, not the model. A tilted or turned head foreshortens one side and skews the widths. Hair over the forehead or jaw hides the exact points the measurements depend on. A lifted or lowered chin shortens or lengthens the apparent face and moves the ratio. Poor lighting that flattens the features can blur where a landmark should sit.

The fixes are simple and worth the effort: face the camera straight on at eye level, pull hair back, keep a neutral expression, and light the face evenly. Get those right and the detector reads the same geometry a tape measure would. When a face still is not found, the troubleshooting guide covers the causes, and the photo guide shows what a clean capture looks like.

Face shape detection, then, is landmark mapping followed by measurement and comparison. The detector places 468 points with MediaPipe, reads the four proportions from them, and classifies the result against the seven outlines, all inside your browser with no upload. The confidence score reports how cleanly your face fits one shape, and the few things that weaken a reading are photo problems rather than model errors. Run a clean, level photo through the detector and the result reflects the same geometry a stylist would measure by hand.