How this face shape detector works
Most face shape tools tell you the answer and nothing about how they got it. This page is the opposite: the full method, so you can judge whether the result deserves your trust, and so you can tell where it is likely to be wrong.
1. Find the landmarks
The first time you analyse a photo, your browser downloads two things: a face landmark model (about 3.6 MB) and the runtime that executes it (about 3 MB, compressed). That is roughly7 MB in total, on the first analysis only — afterwards both come from cache and the tool keeps working offline. It runs locally and returns 478 points on the face — eyes, brows, nose, lips, jaw contour and forehead. Everything after this step is arithmetic on those points.
The model is downloaded to your device; your photo is never sent anywhere. That direction matters, and it is why this tool has no upload endpoint to breach.
2. Take four measurements
| Measurement | Where it is taken |
|---|---|
| Face length (L) | Hairline to the bottom of the chin |
| Cheekbone width (W) | Across the widest point of the cheekbones |
| Forehead width (F) | Across the forehead at its widest point |
| Jaw width (J) | Across the jaw at the jaw angle |
3. Turn them into ratios
The classification uses ratios, never centimetres. This is what makes the result independent of how close you were to the camera and of the photo's resolution — a face photographed from twice as far has half the pixel width but the same proportions.
- L ÷ W — the first fork: long faces (oval, oblong) versus compact ones (round, square).
- F ÷ J — whether the forehead or the jaw is wider.
- W ÷ J — how fast the face narrows from the cheekbones to the jaw.
- F ÷ W — whether the cheekbones are the widest point of the face.
Where the face is widest is what separates the shapes that otherwise look alike, so the last two ratios do most of the work. The guides for each of the six shapes state their expected ranges: face shape guides.
4. Check the head angle first
Ratios cancel out distance, but they do not cancel out rotation. A head turned even ten degrees compresses one side of the face and moves every width measurement. So before classifying anything, the tool estimates the head's roll, yaw and pitch from the landmarks themselves, warns you when the angle is too large, and rotates the landmark set back to level before measuring.
What the result is for
The tool reports two things, and they are not equally trustworthy. Thefour measurements are arithmetic on landmarks: reproducible, and directly interpretable ("your cheekbones are about 27% wider than your jaw"). Thecategory label is a convention applied to those numbers, and it is the weak part.
So read the measurements first. Use the category as an entry point into the styling guides — where a hairstyle adds width, which frames balance which part of the face — and not as a verdict about your face.
What it cannot tell you
- Your hairline is not a landmark. The model estimates the top of the forehead from the visible skin. Hair covering the forehead makes the face read shorter than it is, which pushes the result towards round or square.
- Beards thicken the jaw. A full beard can move a result from heart towards square without any change in bone structure. Re-test clean-shaven if you want the baseline.
- Phone front cameras are wide-angle. Shot from close up, they inflate apparent face width. Stand at arm's length rather than leaning in.
- Borderline faces are genuinely borderline. The six categories divide a continuous space of proportions, and the four ratios have a narrow range — most real faces sit well inside the middle of it. When two categories score close together, the tool shows both rather than pretending the choice is clear.
- "Forehead width" has no single definition. Published guides disagree about where to measure it — at the hairline, midway to the brows, or across the temples. Different choices give different numbers for the same face, which is a large part of why different tools report different shapes. This tool measures at the widest point of the forehead and says so.
- It is not a measurement of attractiveness, and it never scores you. Face shape is a description of proportions; every shape has styling that works with it.
Privacy, concretely
- No upload: the image is read from your device and never sent.
- No storage: no image, no measurement, no facial template is kept anywhere.
- No account, no signup, no email.
- Verifiable: open your browser's developer tools, watch the Network tab, and analyse a photo once the model has loaded. There is nothing to see, because nothing is sent.
You can also use the tool offline after the first load, which is only possible because the computation never leaves the device.
Questions
Is my photo uploaded?
No. The image is decoded, measured and discarded inside your browser tab. The face landmark model is downloaded to your device once; after that, analysing a photo makes no network requests at all. You can confirm this in your browser’s developer tools by watching the Network tab while you analyse a photo.
How accurate is the result?
Lower than we would like, and we would rather say so than invent a number. We evaluated the tool on 200 photos from a public dataset: the category it picked matched the dataset label 30% of the time (top-2, counting the runner-up, 50%). On a five-category problem, chance alone is 20%, so that is a long way from useful — and the labels in that dataset are themselves of unknown quality, since none of the four measured ratios separated the labelled classes at all. What this means in practice: the four measurements the tool reports are reproducible and meaningful (test-retest variation on the same face is small), but the single category label is the weakest part of the output. Read the measurements first, and treat the category as a starting point for the styling guides rather than an answer.
Do you store anything about my face?
No image, no measurement and no facial template. There is no upload endpoint, no account and no analytics event tied to a result. The measurement exists only in the page while you are looking at it.
Do you score attractiveness?
No, deliberately. Face shape describes bone-structure proportions and is useful for styling decisions. Scoring how someone looks is a different product, and not one we are interested in building.