Find your face shape

Seven: oval, round, square, oblong, heart, diamond and triangle. Oval sits widest at the cheek with a gradual curve to the chin. Round and square share their proportions and split on jaw angle. Oblong is a square stretched vertically. Heart and triangle mirror each other, one tapering down and one tapering up, while diamond needs a narrow forehead and jaw under dominant cheekbones.

No. The face mesh is computed locally by Google MediaPipe running inside your browser tab. Your image never travels to our servers, is never written to a database, and clears from memory when you close or reload the page. Nothing is retained after your session ends, and no account is needed to use the tool.

A photo wins on repeatability. Landmark detection places the same anatomical points every time, while a tape slips against soft tissue and depends on how hard you press. A tape wins on depth, since it can follow a curve that a flat image cannot see. For proportional ratios, which is what classification needs, the photo method is the more consistent of the two.

No. Your photograph never leaves your device. Every face tool on The Face Shape Detector runs its analysis inside your own browser using Google MediaPipe, a landmark model the browser downloads once and then executes locally. There is no photo upload endpoint on our servers, so no address exists for an image to be sent to, even by mistake. You can confirm it yourself in about a minute: open your browser's developer tools, switch to the Network tab, and run an analysis. The model file downloads. No image goes anywhere.

Nothing from the tools. Photo analysis and quiz answers alike stay on your device, so color analysis, body shape, body type and Kibbe results are never written anywhere we can read. What The Face Shape Detector does collect is narrow and ordinary: contact form submissions, saved to a database so we can reply; standard server access logs holding IP address, timestamp and browser type; Google Analytics measurement data; and advertising data collected by Google AdSense. The public site has no account system and no login, so no profile of you exists to build.

No, the camera path works the same way. Grant camera permission and your browser hands the video stream to the page, which renders it into a local canvas frame by frame for the model to read. Frames are analyzed and discarded immediately. Nothing is recorded, and nothing is transmitted. Most browsers scope that permission per site and per session, and you can revoke it at any moment from the padlock or camera icon in the address bar. Prefer not to grant it? Use the file picker instead, for an identical result.

The 468 landmarks are numbered points that the MediaPipe Face Mesh model places on a detected face: along the jawline, around each eye, across the brows, down the nose bridge, around the lips and over the cheeks. Each carries an x, a y and an approximate z coordinate. Once the mesh is fitted, geometry becomes arithmetic. Face length is the distance between two points, jaw width the distance between two others, and a face shape classification is a set of ratios between those distances compared against defined ranges.

Because the geometry the model measured genuinely changed between the two shots. A head tilt of a few degrees alters width to height ratios. Side lighting darkens one edge of the jaw and moves the outline. A smile widens the lower face. A close range selfie lens enlarges central features relative to the perimeter. Hair across the forehead hides a measurement the model needs. When two photographs disagree, take a third: square to the camera, neutral expression, even front light, hair back, camera at arm's length or further.

No. There is no app, no plugin, no extension and no account. Everything runs in a standard web browser on a page you have already loaded. The only thing that downloads is the MediaPipe model file, which your browser fetches automatically and caches like any other web asset. Camera access is optional too: you need to grant it only if you choose the camera option rather than picking a file from your device. Nothing is charged for any of this, at any point.

A good photo faces the camera square on, with your eyes level and a neutral, relaxed expression. Use even light from the front. A window during daylight works well, while overhead spotlights and strong side light cast shadows the model may read as an edge. Pull hair away from your forehead and jaw. Hold the camera at arm's length or further to avoid lens distortion. Aim for at least 600 pixels across your face, unfiltered and uncropped. Those five conditions account for most of the variation between two results.

The length-to-width ratio divides total face length, hairline midpoint to chin base, by the widest of the three horizontal measurements. Readings near 1.0 describe a short compact outline. Readings past 1.5 describe an elongated one. It is the first value any classification system computes, and it narrows seven categories down to two or three before any other measurement is applied.

Yes, gradually. Facial fat distribution shifts with age and body weight, which mainly moves apparent cheek and jaw width. Bone structure changes far more slowly, with some remodeling of the gonial angle across decades. A result that changes within weeks is almost always caused by weight, swelling or photo conditions rather than by skeletal change.

Four causes account for almost every disagreement. Hair crossing the temples narrows the detected forehead. A smile widens the cheekbone line. Holding the phone below eye level enlarges the jaw. And a wide-angle front lens exaggerates whatever sits nearest the sensor. Retake the shot at eye level, an arm's length away, hair back, under even frontal light.

Processing happens on the processor inside the device you are holding. When you pick a photo, the browser reads the file into memory, draws it to an in page canvas, and hands those pixels to the MediaPipe model already loaded in the tab. The model returns 468 coordinate points, and The Face Shape Detector's code computes ratios from them. All of it stays inside the browser's own sandbox. No step in the chain sends a network request carrying your image, and closing the tab discards everything.

No, and it would be impossible. Training needs collected images, and none ever reaches us. The landmark model behind the face tools is Google's pre trained MediaPipe Face Mesh, which arrives at your browser already trained, stays unmodified by anything you do, and sends nothing back. No image analyzed here is stored, logged, sampled, queued for review or added to a dataset, ours or anyone else's. That follows from the architecture rather than from a promise somebody has to police.

Broadly yes, with one caveat about the machine rather than the site. Because no image is uploaded, nothing about your face persists on our side. What can persist locally is browser state: a theme preference in local storage, quiz progress in session storage, and analytics or advertising cookies. And if you picked a photo from the machine's own file system, that file is still sitting there. Use a private or incognito window, close the tab when you finish, and delete any file you copied onto the device.

Landmark placement on a clear, front facing, evenly lit photograph is reliable to within a few pixels, and the arithmetic that follows is exact. The uncertainty sits elsewhere. Face shape categories have no agreed scientific boundary, so where one shape ends and the next begins is a judgment encoded in our thresholds rather than a fact about faces. Expect a dependable answer when your proportions sit clearly inside one category, and genuine uncertainty when they sit near a boundary. That is why confidence percentages get published alongside every result.

Because a single answer would misrepresent the certainty. Most faces are not textbook examples of one category; they sit between two. A result of 44 percent oval and 38 percent heart is not the tool failing to decide. It is an accurate description of a face with oval proportions and a heart shaped hairline. Reporting only "oval" would throw away the more useful half of the finding. Read the top two figures together, and where they sit close, treat both sets of styling guidance as relevant to you.

Partly, and not by design. Your first visit needs a connection to load the page and the model file. Once your browser has cached that model, an analysis may still run while you are offline, because the computation itself needs no network at all. We do not guarantee it. Cache behavior varies between browsers, and a cached model can be evicted at any time. Treat offline operation as an occasional side effect rather than a feature you can plan around.

Yes, with a caveat for each. Glasses bias the eye landmarks, so remove them for the eye shape tool and keep them on when you are testing frame styles. A beard changes the visible jaw outline, so a face shape result describes the bearded outline rather than the bone beneath it, which is useful for choosing a haircut and less useful for reading your structure. A headscarf covering the hairline hides the forehead measurement, lowering confidence in face shape but not in eye, nose or lip analysis.