About The Face Shape Detector

Search "what is my face shape" and several hundred million results come back disagreeing with each other. One article measures the jaw at its widest point. The next measures at the gonial angle. A third has you trace your outline on a bathroom mirror with a bar of soap. Even the category list moves: five shapes in one place, seven somewhere else, an eighth invented for the article.

Advice built on that footing is worse than the measurement underneath it. Nobody publishes the thresholds they applied, nobody shows their working, and nobody says how confident the answer is.

The Face Shape Detector began as a fix for the measurement half of the problem. A classification is only useful if it repeats: the same face, shot twice, should land in the same category, and the rule that put it there should sit somewhere you can read it.

The problem in one sentence

Face-shape guidance online contradicts itself because it rests on eyeballing rather than measurement, and because almost nobody states their rule. Stating a rule means naming the landmarks it reads.

Landmarks come from Google MediaPipe Face Landmarker, running as a WebAssembly module inside your browser tab. It places 468 points in three dimensions across the jawline, cheekbones, brow, orbital region, nose, lips and chin. Once the page has loaded, none of that work needs a server round trip.

From those points we derive normalized ratios rather than raw pixel measurements. Pixel distances shift with camera distance, focal length and image resolution, so they carry no meaning across two photographs. Ratios hold steady. Five of them do most of the classifying.

Each ratio meets a threshold drawn from the craniofacial measurement literature that orthodontic and plastic practice already relies on, rather than from a beauty blog's house style. Where sources conflict, the relevant tool page says so instead of quietly picking a winner.

Output is a primary classification with a confidence figure, plus a named secondary shape when ratios sit near a boundary. Faces don't sort cleanly into seven boxes. A face reading 52 percent oval and 41 percent heart gets described that way, and honest description is the first of three mechanisms we'd rather be judged on.

MediaPipe Face Landmarker fits the job because face shape is a landmark problem, not an object problem. Different computer vision tasks need different models, and picking the wrong family makes the answer worse, not better. Here is how the common options compare for reading a face.

So the pipeline runs MediaPipe for the points, then plain geometry on top. YOLO and SAM solve real problems, just not this one. Bolting either on would add tens of megabytes to the download and still leave us measuring the same 468 points underneath. We measure them once, in your browser, and skip the weight.

Trust earned by adjective is worth nothing. Three mechanisms instead.

1. Processing happens on your device

Processing runs entirely inside your browser. Images are never transmitted to, received by, or stored on our servers. There is no photo upload endpoint on this site: not a private one, not an internal one, not one hidden behind a feature flag. Your browser decodes the file you pick, the MediaPipe model reads it in local memory, and every trace disappears when the tab closes.

Quiz and calculator tools behave identically. Color analysis, body shape, body type and Kibbe typing evaluate typed answers in browser JavaScript, and those answers go nowhere.

You don't have to take our word for it. Open developer tools, switch to the Network tab, run an analysis. Model files download once. Nothing carrying your image goes out.

2. The methodology is written down

Methodology pages state which landmarks a tool reads, which ratios it computes and roughly where its boundaries fall. Disagree with a threshold and you can at least see the threshold you're disagreeing with. Our editorial policy sets out the research, review and correction process.

3. We state the limits

Limits get published on the tool itself, never buried in a footnote. Landmark detection degrades under extreme head rotation, heavy shadow, low resolution, obstructive hair and wide-angle distortion from close selfies. Our attractiveness test and age estimator are entertainment built on geometric averages: they report how closely a set of proportions matches a population mean, which says nothing about anyone's worth or health. Naming the people who publish those limits matters as much as publishing them.

Ahtisham ul Haq founded the site and writes the technical material on it. He works in machine learning and deep learning, and his day-to-day subject is convolutional neural networks applied to computer vision. Facial landmark regression, the class of model this site runs on, is the specific problem he works in.

That background is the reason the tools report numbers rather than verdicts. Someone who trains landmark models knows where they break: at extreme head angles, under uneven light, on faces the training distribution underserved. So every result here ships with the ratios behind it and a note on what could have skewed the reading.

You can reach him on LinkedIn. His full profile and every article under his byline sit on his author page.

Every article carries a byline and a last-reviewed date. Tool pages get re-read whenever a model version or a classification threshold changes, and the change is noted.

Editor note: add the registered trading name before launch if the site operates under one.

Accountability also means being blunt about the subjects nobody here is qualified to address.

Medical, dermatological, orthodontic and psychological advice sits outside our scope. A face shape classification describes geometry. It isn't a diagnosis, it isn't a screening tool, and it must never decide anything about your health. Worried about your jaw, skin or teeth? See a clinician.

Selling personal information sits outside it too. The public site runs no account system and no login, so there's no profile for anyone to buy. The only personal data held here is whatever you type into the contact form, plus the standard analytics and server logs described in our privacy policy.

Nor do we claim a classification is objectively correct in some absolute sense. Each one is the output of a stated rule applied to measured ratios. A different rule returns a different answer, which is precisely why ours gets published. Publishing costs us nothing, because the revenue arrives elsewhere.

Funding comes from display advertising served through Google AdSense. Nobody pays us for a particular result, and advertisers get no input into editorial content, rankings or methodology. Ad units stay visually distinct from article copy. Our editorial policy details that separation, and the privacy policy covers the data advertising involves.

The same revenue pays for the inbox where corrections land.

Errors are worth reporting. A threshold that looks wrong, a broken tool, a factual mistake, an unsupported claim: email [email protected] with the page URL and, where you have one, the source you believe we contradicted. Reports get acknowledged within five working days, and confirmed factual errors are corrected within ten working days of confirmation.

General questions, press inquiries and permission requests reach the same address, or use the contact form at https://thefaceshapedetector.net.

The Face Shape Detector exists because facial proportion can be measured instead of eyeballed, and because the rule behind any measurement should be readable. Fourteen browser-based tools apply that idea, from face shape detection through eye, nose and lip work to color, body and Kibbe typing. Google MediaPipe Face Landmarker supplies the 468 three-dimensional points; normalized ratios follow, thresholds decide, and a confidence figure reports how close the call was.

Three mechanisms carry the result: analysis on the hardware in your hand with no upload endpoint anywhere, published methodology, published limits. Briefed writers draft, a second reader checks. Health questions belong to clinicians. Google AdSense keeps every tool free and touches nothing editorial.

Find something wrong here, and [email protected] is the address.