What a Face Shape Detector Actually Decides

A face shape detector reduces your face to four straight lines — forehead width, cheekbone width, jaw width, and length from hairline to chin — then hands you whichever label those four numbers sit closest to. That is the whole mechanism. Browser tool working off a photo, or you alone with a tape measure: either way the output is a ranking of four numbers, and the label is just the name of that ranking. So the useful question is not which detector is best. It is which of the four numbers the tool actually measured, and where it drew the cut-off.

Neither of those is standardised. That is the part that goes unsaid.

The cut-offs disagree, and you can check the arithmetic

Take length divided by cheekbone width, the one ratio nearly every tool leans on hardest.

• goldenratioface.net calls a face rectangle at 1.50 or above, with a width variance of about 15 percent.
• faceshapedetector.app puts oblong above 1.5, oval at 1.3–1.5, and a joint oval/square band at 1.1–1.3.
• faceshapetool.com does not reach oblong until 1.6 or higher.
• LoopedInLooks describes oval as a length 40 to 60 percent longer than the cheekbones, which is 1.40 to 1.60.

Now run a ratio of 1.55 through all four. Two of them hand you oblong. Two of them still call you oval. Same face, same number, two answers, and neither is a malfunction. They drew the line in different places, that is all.

There is one place they converge. Round ends at ten percent: faceshapedetector.app cuts round off below 1.1, and LoopedInLooks defines round as a length within 10 percent of cheekbone width. If your ratio comes out at 1.05, every route agrees with every other. Between 1.10 and 1.40 the published bands thin out badly, and that stretch is where a great many faces sit.

The jaw number is not even the same number

This is worse than an argument about thresholds, because it corrupts the input before any threshold gets applied.

One method runs the tape from the base of one ear, down along the bone, to the tip of the chin, and then doubles it. The other asks for the flat width from one jaw corner to the other, explicitly not doubled. LoopedInLooks uses the first. faceshapedetector.app uses the second.

These are not comparable quantities. The first traces a curve around the underside of the face; the second is a chord straight across it. Work it through on a real set of measurements — forehead 6.5 inches, cheekbones 7, half-jaw 5, length 8. Doubling the jaw gives 10. Ten is now the largest of the four by a wide margin. Under any rule that assigns the label to the widest line, a jaw of 10 against cheekbones of 7 reads as triangular or pear: bottom-heavy, jaw dominant. The face those numbers came from is not bottom-heavy. The doubling instruction did that, not the bone.

If a doubling method told you triangle or pear, measure the flat jaw width and run it again.

The forehead line moves depending on who is asking

Four different heights, each presented as "the forehead measurement":

• the widest point just below the hairline
• midway between the hairline and the eyebrows
• across from the peak of one eyebrow arch to the peak of the other
• hairline to hairline, at the point where the forehead is tallest

On most foreheads these are not the same span. The brow-arch line sits lower and comes out narrower than the hairline line. And forehead width is exactly what separates heart from diamond from square, so choosing the lower landmark can move you a full category. No one publishes which height is correct. There is nothing to publish.

The phone filter route closed on 14 January 2025

Searching "face shape" inside Instagram's effect gallery and cycling the analysis filters used to be the fastest answer going. It does not work now. Meta discontinued Meta Spark, and third-party effects stopped being available on Instagram and Facebook effective 14 January 2025 — the face-shape analysers were third-party effects. Guides still walking you through that path are describing something that has been gone for over a year.

Worth knowing before you hunt for a replacement: it was unreliable while it existed. The same face, held still, cycled through several of those filters, came back rectangle, then square, then triangle, then oval. The failure was visible on screen. The detection box did not cover the whole face, so the filter was classifying a crop, not a head.

Seven shapes, or eight, or four

The count depends entirely on where you looked. faceshapedetector.app and faceshapetool.com both settle on seven: oval, round, square, heart, diamond, oblong, triangle. goldenratioface.net runs eight. Another well-travelled set of seven drops triangle and adds pear, where the jaw is the widest line and the forehead the narrowest — and pear gets left out of most lists despite being genuinely distinct.

Then there is a four-shape system that throws the tape measure out completely. It judges the outline only, off two features — how prominent the cheekbones are, how prominent the jaw is. That yields oval (neither), diamond (cheekbones only), square (jaw only), and hourglass (both). Hourglass has a tell you can spot without measuring anything: the natural shadow between cheekbone and jaw that most people have to fake with contour. In this system round is not a shape at all. It is a short oval. Heart is not a shape either; it sits between diamond and hourglass. The four are arranged as a loop rather than a list, on the reasoning that most people land between two neighbours instead of inside one.

That is not a quibble about naming. It means round and heart exist or do not exist depending on which page you opened.

Your label can shift without your bones shifting

Weight shows on the face early for a lot of people, and a fleshier face reads rounder. The drift is concrete enough to describe: a face with a visible jaw angle and prominent cheekbones reads square while it is lean, then reads round as it fills out, the jaw angle no longer obvious. Both readings were accurate when they were made. What follows from that is blunt — if you are carrying extra weight, a round result may be describing the weight rather than the skull.

Resist one more thing: the notion that oval is the safe result that suits everything. An oval reads as oblong the moment you wear the hair long, straight and centre-parted, because that emphasises length. Thick bangs straight across do the opposite and push it toward round or square. The label is a starting position, not a property you own.

A sheet you can fill in

Every row below uses the flat jaw width — corner to corner, not doubled — and a face length taken from the centre of the hairline straight down to the tip of the chin. Pick those two conventions and hold to them, because mixing them with the doubled-curve method is exactly what manufactures a false triangle. The ratio column names the tool each threshold came from; where no tool publishes a number, the cell reads Not published instead of a range I would have had to invent.

Two things worth reading off it. Oval and Rectangle carry two ranges each because the tools genuinely disagree at those boundaries, and the Hourglass row is nearly empty because that shape gets assigned from the outline rather than from a tape. Nothing measurable in these rule sets separates Triangle from Pear either — both are jaw-widest, forehead-narrowest, longer than wide — so that distinction is marked Not published as well.

Downloadable copy: face-shape-measurement-sheet.csv, the same nine rows plus a blank block at the bottom for your own four numbers. It opens in Excel, Numbers or Google Sheets.

What to do with the four numbers

Measure the flat jaw width, not the doubled curve. Write down the ratio of length to cheekbone width, and treat that number rather than the label as your actual result. Under 1.10, you are in the single zone where everything agrees. Between 1.30 and 1.60, expect different tools to name you differently, and go with the one that matches what you can see in a mirror. And note what is missing: none of the four rule sets above publishes an error rate or a confidence figure beside its answer. Two of them advertise 468 landmark points — the same count as Google's MediaPipe Face Mesh model — and then collapse all 468 into four straight lines and one ratio.