Is a photo more accurate than a tape measure?

Not necessarily, and they measure slightly different things. A tape goes across the widest point it can find; a mesh takes the distance between two fixed points. The mesh is more repeatable and less prone to a shaky hand, and a tape is less prone to lens distortion and head tilt. This site converts the mesh reading onto the same scale as the tape figures so the two routes can be compared, and it says so on the result.

Why did you publish a failure of your own tool?

Because the alternative is another page claiming accuracy without evidence, which is what the rest of this category already offers. A measured failure with its cause and its fix is more useful to somebody deciding whether to trust any of these tools than a number with no method attached.

Does head tilt or camera distance change the result?

Tilt does, which is why the tool refuses a face turned too far rather than measuring it anyway. Distance does not, because every width is divided by cheekbone width, so the result is a set of proportions rather than sizes. A very close phone camera can still stretch the centre of the face, so an arm's length is better than very close.

Should I trust the result if it says the call was close?

Treat it as two candidates rather than one answer. A close call means the proportions genuinely sit near a boundary, and in that case both sets of styling advice are worth reading. It is not the tool hedging. It is the tool declining to invent a precision the measurement does not support.

How accurate is a face shape detector? One tool answered heart for 27 faces out of 32, and the reason was not the AI

The tool on this site was wrong for a period, in a specific and measurable way. The record of how that was found is more useful than another claim of accuracy, and it gives you a test you can run on any tool in about two minutes.

Last reviewed 11 August 20268 min readSravya

Editorial portrait with the face outline clearly visible for measurement

The figures

27 of 32
Real faces this tool once called heart: 0.7%
Variation in the broken measurement: 7 of 7
Reference faces correct after the fix

The short version

A face shape detector can be accurate about geometry and cannot be accurate about a category, because the categories are conventions rather than facts. What separates a usable tool from a guess is whether it shows the measurements, publishes the thresholds it compares them against, and says when the call was close. Almost none do. The failure described on this page was not a failure of the model: the face mesh reported the landmarks correctly every time. It was a failure of which two landmarks were being measured between, which is exactly the kind of fault that a tool showing only a one word answer can carry indefinitely without anybody noticing.

01What the fault actually was

The photo route places a 478 point mesh on a face and takes four distances from it: face length, cheekbone width, forehead width and jaw width. Three ratios come out of those, and the ratios are compared against published thresholds. The mesh was never the problem. The problem was the pair of points chosen for forehead width.

The pair in use sat on the widest part of the temporal contour, which is close to the widest part of the head rather than the forehead. Measured across 32 external portraits, that distance came back between 0.994 and 1.029 of cheekbone width for every single face. A standard deviation of 0.7 percent across 32 different people is not a measurement. It is a constant wearing the label of a measurement.

The corrected pair sits on the forehead itself, roughly halfway between the brows and the hairline. Across the same 32 portraits it averages 0.851 with a standard deviation of 2.8 percent, which carries four times the signal and also agrees with published forehead to cheekbone breadth. The tool now spreads those 32 faces across all seven shapes instead of putting 84 percent of them in one.

02Why nobody noticed, including the tests

This is the part worth generalising, because it is not specific to this site. The tool had a test suite. Every test passed throughout. Not one of them caught it.

The tests started from ratios, not from images. They checked that a set of numbers classified correctly, which it always did. The fault was upstream, in how the numbers were produced from a photograph.
The tool never claimed false confidence. It reported close calls honestly and showed the runner up, which made the wrong answers look reasonable rather than broken.
A one word answer is unfalsifiable. Told your face is heart shaped, most people have no way to check, and a reader who disagrees assumes the tool knows better.

A tool that shows only its conclusion cannot be audited by the person using it, which means it can be wrong for as long as nobody happens to look.
The reason every result on this site prints the figures it used

03The test you can run on any face shape tool

Two minutes, no equipment, and it works on any tool including this one.

1Run several clearly different faces through it. Photographs of people with obviously different proportions, not four photographs of yourself.
2If most of them come back as the same shape, the tool is measuring something that does not vary between faces. That is the failure described above and it is invisible from a single result.
3Ask whether the tool shows you the numbers. If it does not, you cannot check the answer and neither can anyone else.
4Ask whether it publishes the thresholds those numbers are compared against. A threshold you cannot see is an opinion presented as a measurement.
5Photograph the same face twice, once in daylight and once under a lamp. A face shape result should not move. If it does, the tool is reading the lighting rather than the geometry.

04What accuracy can reasonably mean here

Three separate claims get bundled into the word accurate, and only two of them can be checked.

Three claims, and which can be tested

Claim	Can it be checked?	How
The measurements are right	Yes	Compare the tool's figures against the same widths taken with a tape measure. They should agree to within a few percent.
The same face gives the same answer	Yes	Photograph twice, in different light, at the same angle. The result should not change.
The label is correct	No	There is no ground truth. The categories are conventions and the boundaries are editorial, which is why this site publishes the thresholds it uses.

A tool that will not let you test the first two is asking to be trusted on the third, which is the only one nobody can settle.

The seven reference photographs used on this site were deliberately held out of the calibration, so they test it rather than confirm it. Before the fix, one of the seven was identified correctly. After it, all seven are. That is a real number about a small sample, which is a more honest thing to publish than a percentage with no method behind it.

Common questions

Sources

Every figure quoted above is listed here with what it was used for and when the source was last checked. Anything not attributed to a source is this site’s own reading and is labelled as such where it appears.

Google
MediaPipe Face Landmarker for the web
The 478 point face mesh used by the photo routes, and the fact that it runs locally in the browser rather than on a server.
Last checked 11 August 2026

Check it against your own measurements

The result prints the three ratios it used, how strongly every one of the seven shapes matched, and whether the runner up was close. That is what makes the test on this page possible to run against it.

Measure it and check the working