The AI Model Behind Face Shape Detection

MediaPipe Face Mesh is Google's face-landmark model, part of the open MediaPipe framework for on-device machine learning. It was trained on a large, annotated set of face images to predict the position of a fixed set of landmark points on any new face, and it returns that full set in a fraction of a second.

The model is built for exactly this kind of use: real-time, on-device face geometry, without a server. Google designed it to run on phones and in browsers, so it is compact and fast rather than huge and cloud-bound. That is why a face tool can use it locally at all. The site loads the model through MediaPipe's Face Landmarker task, which wraps the mesh model with the code that finds the face first and then places the points on it.

The model outputs 468 points, each a coordinate on the face, together forming a dense mesh over the brow, eyes, nose, cheeks, mouth and jaw. Each point has a consistent identity: the same index always marks, say, the center of the chin or the outer corner of an eye, across every face the model reads.

That consistency is what makes measurement possible. The site does not hunt for features from scratch; it reads specific landmark indices to find the hairline center, the widest cheek points and the mandible angle, then measures the distances between them. The model also infers a rough depth for each point, a third dimension, which helps it place landmarks correctly when a face is turned slightly. With 468 points, no single landmark has to be perfect, since each measurement reads from a cluster and a small error in one place barely moves the result.

Running a machine learning model in a web page is possible because MediaPipe ships two pieces the browser downloads: a WebAssembly runtime and a model file of roughly 4 megabytes. WebAssembly lets the browser execute the model's math at near-native speed, and the model file holds the trained weights. Together they run the face mesh locally, using the device's processor.

The download happens once. On your first visit the browser fetches both pieces and caches them, so every later analysis skips the download and starts immediately. There is no server call after that one fetch, which is why results appear in seconds and why the tool works the same across modern browsers. The model runs on the CPU by default and can use the graphics processor where the browser allows, but either way the computation stays on your device. The mobile page covers how the same download behaves on a phone.

It is worth being precise about the division of labor. The model decides where the landmarks are. It does not decide your face shape. Once the points are placed, the site computes the four measurements, forehead, cheekbone and jaw width against face length, and compares them to the seven outlines using the same tables published on the shape pages.

That means the shape classification is transparent geometry, not a hidden judgment. You can check it: run the measurements the detector shows against the face shapes hub and the same logic applies. The model's job ends at the mesh, and everything downstream is arithmetic you could do by hand with a ruler. Keeping the learned part and the rule-based part separate is what lets the tool explain its own reasoning rather than asking you to trust a score.

Running the model locally is a privacy decision as much as a technical one. Because MediaPipe executes in the browser, the photo never has to travel to a server to be analyzed, so there is no upload endpoint, no transmission and no storage of your image anywhere. Closing the tab discards everything.

There is a practical payoff too. On-device analysis has no network round trip after the one-time model download, so it is fast and works without an account or a queue. It also means the tool keeps working the same regardless of server load, since there is no server in the loop. This on-device design is a load-bearing part of how the whole site is built, and it is the reason the privacy claim holds rather than being a promise you have to take on faith.

The face-shape detector is not the only tool built on this model. The nose shape and eye shape readers use the same 468-point mesh, reading different landmark clusters to measure the nose bridge or the eye corners, so they inherit the same on-device, no-upload design. The skin tone analyzer reads pixel colors from the photo rather than the mesh, but still runs locally.

Two tools you might expect are deliberately absent: an age estimator and a celebrity matcher are not implemented, and the site reports them as unavailable rather than returning a made-up answer. That honesty is intentional. A tool that cannot do something well should say so, not guess. The models that are here, the face mesh and the pixel reader, do defined jobs the site can stand behind, which is why the results come with measurements you can check.

The AI model behind face shape detection is Google MediaPipe Face Mesh, a trained model that locates 468 facial landmarks and nothing more. It runs in your browser through a WebAssembly runtime and a roughly 4-megabyte model file downloaded once, so the photo never leaves your device. The model places the points; the site's own geometry turns them into one of seven shapes, a split that keeps the reasoning transparent and the analysis local. The same mesh powers the nose and eye readers, while the age and celebrity features stay unbuilt rather than faked. See how the pieces fit on the how it works page, then run a photo through the detector.