Round vs Square Face Shape: The One Check You Are Allowed to Leave Blank

The Strip Where Both Labels Stay Possible

Start with what a tape gives you, because that is the half a reader can check at home. This site's rule set puts a round face's forehead at 0.88 of the cheekbone width and a square face's at 0.95; the jaw at 0.85 and at 0.95. Each of those centres has a published tolerance, and a band is one tolerance either side of the centre. Where the two bands overlap is ground on which your measurement is consistent with both labels, and that ground is not narrow.

The table below is that overlapping ground, multiplied out at the three cheekbone widths this site works from: the narrowest and widest adult face in its published band, and its working line. Every cell is one multiplication of a published band edge by the row's width, and one division by 25.4. Any row can be redone on the spot with your own cheekbone width in place of the printed one, and the arithmetic is one step, not a model.

Read the fourth column twice, because it is the odd one. The forehead and jaw strips grow with your face width; the angle strip does not. Degrees are not a proportion of anything here, so the same three numbers sit at every width in the table.

Centres and tolerances are imported from this site's published rule set, on the soft-tape scale: forehead 0.88 against 0.95, tolerance 0.081 either way; jaw 0.85 against 0.95, tolerance 0.12; jaw angle 132° against 108°, tolerance 7.66°. Each millimetre cell is a published band edge multiplied by the row's cheekbone width and rounded to one decimal; each inch cell is that same unrounded product divided by 25.4 and rounded separately, so an inch can land a hundredth away from dividing the millimetres printed beside it — the rounding convention is stated on this site's neckline guide for a round face and is followed rather than restated here. One tolerance is a chosen band width, not a natural boundary, and widening it changes what follows: at two tolerances the forehead strip becomes 0.254 wide and the jaw strip 0.378 (34.5 mm and 51.5 mm at the 136 mm line), and the angle strip closes from a gap into an overlap running 116.7° to 123.3°, 6.6° wide. What does not depend on the band width is which rows touch and which do not. 128.2 and 148.8 mm are the ends of the adult cheekbone band this site works from and 136 mm is its working line; the dataset behind that band is named where this site publishes it and is not reprinted here.

What the Published Answers Say, Counted with One Ruler

Ten pages answer this question at the top of ordinary search. Six of them gave up their text to be read; two sat behind a bot wall that returned an interstitial instead of an article, one is an archived discussion whose visible text has been removed and whose first screen held 1,132 characters in total, and one is a short-video page that publishes a caption and no subtitles. Those four are not evidence for anything, they are not rows in any table below, and nothing on this page rests on what they might have said.

The six readable pages were then counted with one yardstick: how many tables, how many millimetres, how many degree marks, how long the body. This site's own comparison page for a different pair was counted the same way so the numbers mean the same thing on both sides, and it returns three tables, two millimetre strings and sixteen degree marks. Of the six external pages, six out of six carry zero degree marks. Five of the six carry zero millimetres. One carries thirteen tables and not a single measurement in any cell of them — they are makeup and hairline comparison tables. In the whole set there is no number attached to the border between these two faces.

What they do say, in their own words. A makeup artist's blog, dated 2020 and revised 2025, the longest page in the set: both shapes are `a short & wide face shape`, and `the only thing that is going to be different is the angle of the jawline & hairline`, the round one curving, the square one `more straight on the sides of the face`. It names the corner, and gives it no number; the word ratio appears zero times on that page. A clinic's facial-geometry assessment from 2026 goes further and gets closer: `A face can be as wide as it is long and still not be round if that jaw turns a corner — then it is square`. That is a real criterion. It has no degrees on it, and its two tables contain no millimetres. A cosmetics brand's how-to page gives you four tape steps and lands on `If all proportions are nearly equal, it's probably square or round` — four measurements ending in a tie, with no band and no cut-off.

The other three are thinner again. An eyewear brand's blog mentions a tape measure fifteen times and never attaches a number to a criterion. A booking platform's style guide carries exactly one sentence in which both shape names appear in the same place: `Round faces have similar traits, but they have more curves than square faces`. A barbershop's assessment page lets the two words co-occur only inside lists. And the generated answer block at the top of the results screen, which rewrites the first of those pages for you, contains one number in the whole of its comparison: a 1:1 ratio of length to width. Nothing in it puts a measurement on the corner it tells you to look at.

This site publishes the five checks, the seven rows, the centres and the tolerances, with the calibration and the two measurement scales named: Read the full rule set ↓

Two of the Five Checks Never Move

Put the two rows side by side on all five checks and two of them are the identical printed number: the face length divided by cheekbone width is 1.05 on both rows, and the chin width divided by jaw width is 0.55 on both. This is not a rounding accident in two separate calculations; the rule set holds the same cell for both labels, on purpose. Someone already counted how many checks print the identical value for this pair and put the count in a table on this site's diamond-versus-heart page, so this page will not run the count a second time. Read that row and come back.

The two width checks behave differently, and the first table on this page is what they do. Neither pair of bands is disjoint. On the forehead row, the round band runs 0.80 to 0.96 and the square band 0.87 to 1.03, so each centre sits inside the other's band and the shared ground is 0.09 of your cheekbone width wide. On the jaw row the same thing happens at a different scale: 0.73 to 0.97 against 0.83 to 1.07, shared ground 0.14 of the cheekbone width, and again each centre falls inside the other row's band.

Say what that means before it gets over-read. A forehead width and a jaw width taken off your own face can be entirely consistent with both labels at once. That is not a failure of your tape; it is what those two rows of the published table say. The guides that answer this question with soft against straight are not wrong about the overlap — they simply never mention it, because mentioning it would end the sentence.

Two of the five checks identical, two of the five overlapping, one of the five carrying a real gap. That is the arithmetic of this pair, and it points at exactly one row for whatever follows. The millimetre version of the two centres, and the size of the difference between them expressed as a fraction of the spread inside a single shape, are worked out on this site's neckline guide for a round face. This page multiplies out only the overlapping ground, because that is the quantity nobody has printed yet, and it is the one a reader needs in order to know when their own two widths have stopped being able to tell these rows apart.

The Angle Row, in Degrees

On the tape scale this site's rule set puts a round face's corner at 132° and a square face's at 108°. That is the largest gap the two rows have on any check: 24°, which is 3.13 times the tolerance carried inside a single shape on that row. Read the same two rows on the photograph scale the detector works in and they print 142 and 127.9 against a tolerance of 4.5 — different numbers, different ruler, and the identical quotient of 3.13. That is not a coincidence to be explained. The two scales are joined by one straight-line map applied to centres and tolerances together, so a gap expressed as a count of tolerances is a property of the rule set rather than of whichever ruler you happen to be holding.

Unlike the two width rows, these bands do not touch. The square row's ceiling is 115.66° and the round row's floor is 124.34°, so between them sits a strip 8.68° wide that neither label occupies. It is the only such strip in this pair's five checks, and it is the fourth column of the first table: the same three numbers at every cheekbone width, because degrees are not a proportion of anything here. Sensitivity, since one tolerance is a chosen band width rather than a natural boundary: widen both bands to two tolerances and the strip stops being empty, becoming an overlap that runs 116.7° to 123.3° and is 6.6° wide. What survives either band width is that this row carries the pair's widest gap.

There is a second fact about this column that is easy to miss and worth more. Take all seven rows in the published table and read down the angle column. It runs from 108 to 132, and this pair owns both ends of it. The other five rows — oval, oblong, heart, diamond, triangle — sit between 118.04 and 121.96, a span of 3.9° shared by five faces. So the rule set writes these two rows furthest apart at the angle, and it is also the check this site's own photographs separate least cleanly, which is why the classifier scores all five readings at once and a result page returns a margin rather than a verdict. Both halves of that sentence are on the record at /about; the pairing of them is what this section is for.

Say what the row is not, before the next section. The two centres on the tape scale come from a styling convention that predates any measurement programme, not from photographs this site took. It is the one field of the five that a reader with a soft tape and a mirror cannot be expected to produce, and it is the field where the printed table and the dental literature are, as the next section sets out, not describing the same thing in the same units.

What the Same Corner Measures in a Dental Sample

Somebody has measured the corner of the jaw at scale, in clinical conditions, and published it. An orthodontic study of 166 pre-treatment lateral cephalograms, patients aged 17 to 30 at a dental college in Kathmandu, radiographs collected between 8 December 2020 and 8 February 2021, reports the gonial angle — the corner where the body of the mandible meets its rising ramus, which is the anatomical corner a styling page means when it says the jaw turns a corner instead of curving.

Its published means: 132.84 ± 3.70° in the hyperdivergent group, 124.06 ± 3.88° in the normodivergent group, 119.94 ± 5.57° in the hypodivergent group. Split by sex the six means run from 119.14 ± 6.42° to 133.07 ± 3.28°. Within-group spreads of 3.3° to 6.4°, in a sample of 166, against the 7.66° tolerance this site carries on the same column for the tape scale.

Now put those three means against this site's two bands, which is the comparison nobody has printed. The round centre, 132°, sits inside the round band and within a degree of the hyperdivergent mean; that is a coincidence of numbers, not a confirmation, because the study never mentions face shapes. The other two means, 124.06° and 119.94°, land inside the corridor this site's two bands leave empty. And none of the six subgroup means lands inside the square band at all, whose ceiling is 115.66° — this page's own subtraction puts the square centre 11.9° below the lowest of the three group means, about two of that group's own standard deviations down.

What that does to the argument. It does not settle which of the two labels a face wears: the study sorted its patients by vertical growth pattern, not by outline, and no participant in it was ever called round or square, so it is not evidence about this pair, it is evidence about the range of the human gonion. It also measures bone on a radiograph rather than the soft-tissue line a mirror shows, which is a different quantity wearing a similar name. What it does knock down is a stronger sentence this page could have written and did not: that nobody has numbers for the jaw corner. Plenty do; they are in orthodontic journals, they are tight, and they are not arranged around the two words you are searching.

The reason it is worth reading alongside the rule set is the size of the disagreement about where normal sits. A styling table puts one of these two faces at 108°. A cephalogram study of 166 people finds no group mean below 119.94°, and a self-published reference chart of facial proportions from August 2026, which reproduces the same study and is the kind of thing that gets screenshotted, puts the place where adult averages land at around 120° to 125° and warns in its own note that the values vary with population and with measurement method. If the corner were the whole answer, the two literatures would be answering the same question. They are not, and a reader who is being handed a jaw angle as the answer to this question should know that the number crosses a field boundary on the way to them.

Provenance discipline, stated once because it matters here: the three means above were read from the study's own published abstract rather than from the 2026 chart that quotes it, and the two were compared the same day. The chart's rounded 124.1 ± 3.9 is the abstract's 124.06 ± 3.88 at one decimal, so the chain is consistent; this page still prints the primary numbers.

The Blank Field, and What Fills It

The calculator that fills this field in for you, with the estimated corner marked on the result: Face shape calculator ↓

One of the five checks is not asked for. On this site's calculator the fifth input reads `Jaw angle (optional)`, and the four numbers the form requires are the ones a soft tape can reach: face length, forehead width, cheekbone width, jaw width. The corner is optional because it is the quantity nobody can lay a tape along on their own face. The form will accept anything between 80° and 160° if you insist on typing it, which is a wide enough window that typing is not the same act as measuring. So the rule set carries a check its own input screen declines to require, and this is the section about what happens in that gap.

What fills it is not a placeholder. Leave the field blank and the site runs your four widths through the classifier once with the angle row carrying no weight at all, then takes the seven published angle centres and averages them, each weighted by how strongly your four numbers already look like that row. The output is a corner for your face that the rule set never printed for anybody. Worked through the whole pipeline at the 136 mm cheekbone line with a face length of 153 mm, three sets of widths give three filled-in values. Forehead 119.8 mm with jaw 115.5 mm comes out at 123.83°. Forehead 129.3 mm with jaw 129.3 mm comes out at 116.00°. Forehead 124.5 mm with jaw 122.4 mm comes out at 120.08°. All three sit inside the corridor the two published bands leave empty, whose edges are 115.66° and 124.34°: the square-ish face lands 0.34° above the ceiling of the square band, the round-ish one 0.51° below the floor of the round band, and the middle one near the centre of the strip.

That is the honest shape of an estimate. The estimator is not deciding which of these two rows you are on and then printing that row's corner; it is averaging seven centres weighted by your own widths, and the average of a set that contains both 108 and 132 lands where neither of them lives. A reader who types nothing gets a number, and the number is a summary of what the four widths already said rather than a fifth measurement pretending to be one.

Typing the corner moves the printed confidence, and the movement is worth looking at as arithmetic rather than as advice. Beside the round-ish widths, 132° takes the round row from 0.6977 to 0.9514, while the runner-up falls from 0.0921 to 0.0266. Beside the square-ish widths, 108° takes square from 0.7816 — which is printed as 0.70, because an estimated corner puts a ceiling of 0.70 on the confidence — to 0.9720, uncapped. Type 120°, the middle of the corridor, and both readings soften: 0.4005 on the round-ish widths with the square row at 0.1967 behind it, 0.5085 on the square-ish widths with round at 0.2285 behind it. The weight this site carries on the angle row is 0.85, and 1.275 when the value is typed rather than estimated; the 1.5 multiplier between the two is a stated assumption, not a fitted number, because the calibration set contains no person who both self-measured that corner and had it estimated, so there is nothing there to fit it against. None of the figures above is a claim that the corner tells you which row you are on. The classifier scores all five readings at once and a result page returns a margin, which is the reason this site gives on its own two pillar pages for these shapes and in the footnote to the published table.

The result page marks the field when it was filled in for you, and the marking is on the page rather than buried in a tooltip; the two pillar pages and the two result pages for these shapes say in plain sight that this row is the one the site's own photographs separate least cleanly. That disclosure is the point of the optional field. One more reading, because it is cheap and it matters to anyone working in inches: the same face entered as 5.354 in of cheekbone, 6.024 of length, 4.717 of forehead and 4.547 of jaw returns 0.6968 where the centimetre entry returned 0.6977, and the same 123.83° corner. Choosing a unit moves the fourth decimal place, not the answer.

What the form tells you to do about all of this, verbatim: `Hard to measure on yourself. Leave it blank and we estimate it from your four numbers — and say so on the result.` The arithmetic above is what that sentence is short for.

A Face the Calculator Will Not Accept

The form screens for typos before it classifies. A face length under 15 cm or over 30 is refused, and so is any of the three widths outside 8 to 20 cm, and a typed angle outside 80° to 160°. These are guards against a tape read in the wrong unit or a digit dropped, not claims about heads; nobody is asserting that a 149 mm face does not exist.

Then put this page's own table through that screen. Take the round row's three published centres, multiply them by the 136 mm working line and enter the product as a face: forehead 119.8 mm, cheekbone 136.0 mm, jaw 115.5 mm, and a face length of 1.05 × 136 mm = 142.8 mm (5.62 in). The form refuses it, at the length field, with the words `That looks outside the usual range — double-check the number.` The square row gives the same result, because the two rows share the length cell: same 1.05, same 142.8 mm, same refusal.

Work the boundary once, because it is a division and nothing else. 150 mm divided by 1.05 is 142.9 mm. So a face built on either row's own length centre is admissible only at cheekbone widths of 142.9 mm and up — the top 5.9 mm of the adult band this site works from, which runs 128.2 to 148.8 mm. At the 136 mm working line, the part of the round row's own length band that the form will accept begins at a ratio of 1.103 and runs to 1.232; the published centre of 1.05 sits below the line where the tool stops listening.

What to take from that, stated carefully. The table on this page is a scale, not a face. A centre in a ratio table is a convention about a proportion, and the plausibility window is a guard on an input box, and where the two meet a printed centre can be unreachable on the very form that publishes the row. That is not an error to be ironed out by rounding one of them; the two numbers were written for different jobs. For a reader it has one practical consequence: if the calculator refuses one of your four numbers, the suspect is the number in the field, usually a length taken along the curve of the face instead of straight down, and not your face.

Eight Years of Answers to This Question, in Date Order

The rows below are the positions found for this question, oldest first, each with the date printed on its own page. Nothing outside this site's rows has a number against the border. Nothing outside this site's rows has a band either.

Identities are given by type, not by name or link, for every row that is not this site. Dates are the publication date printed on the page, with the revision date where the page carries one. The orthodontic row carries a year rather than a printed day, so it sits at the head of 2021 and no order inside that year is claimed. Not in this table, and not evidence for any sentence on this page: two pages in the top ten with no date anywhere on them, two behind a bot wall that returned an interstitial instead of an article, one archived discussion whose visible text has been removed, and one short-video page with a caption and no subtitles. The generated answer block at the head of the results for this question is a rewrite of the first row above, carries one number in its entire comparison, and is not counted as a source.

What Moved, and What Did Not

Down the middle column of that table, over six years and seven months, the format has not moved at all. Every external row is a description of an outline: curves against straight sides, a corner below the ear, a flat chin line, a hairline. Two of them name the corner as the thing that differs. Not one of them puts a number on it, and not one of them tells a reader what to do when their own face is consistent with both descriptions.

What did move is the honesty of the hedge, and it moved slowly. The 2020 page says the two differ at the jaw angle and leaves it there. The 2021 how-to page runs four tape measurements and arrives at a tie without calling it one. The 2026 clinic page gets closest to a test a reader could apply and still declines to attach degrees to it. Meanwhile the 2021 orthodontic study in the second row measured the same corner in 166 people to two decimal places, for reasons that have nothing to do with face shapes, and that precision has never been carried across to the question people actually ask.

So the gap this page sits in is not a missing opinion. There are six dated opinions and they all say roughly the same thing. The missing part is the arithmetic: one published table of measured angles that is about a different question, one self-published chart that reproduces it and rounds it, and no table anywhere that says where the two labels stop being separable by a tape. The three rows at the foot of that table are this site's attempt to be the row with the numbers in it, and the columns above them are what it is trying to be different from.

If Your Numbers Land in the Corridor

Take the four numbers the form asks for and compare them as themselves before dividing anything. Pull the hair back off the forehead so a fringe is not lowering the reading, hold your head level, and take the forehead across its widest point, the cheekbones from the top of one to the other just under the eyes, the jaw corner to corner straight across, and the length from the hairline to the chin tip. Write the four down. The millimetre cells in this page's first table are the shared ground at three widths; if your forehead and jaw both fall inside their column, your tape has stopped being able to tell these two rows apart, and that is a fact about the bands rather than about your hands.

Then do the one thing a mirror can do better than a calculator. Feel along the side of your face from the corner under the ear down to the chin. A distinct corner and a flat lower edge put you on the square row's side of the description; a continuous curve from ear to chin puts you on the round row's side. That is the sentence every page in the table above has been writing for six years without a number on it, and it is worth doing because it is not a measurement and therefore has no tolerance to argue about.

If the two still disagree, do not average the two pages of advice. This site's hairstyle guide for a round face has a section written for exactly this situation, titled `If Your Numbers Sit Between Round and Square`, and it says why: the two shapes take opposite instructions about the top of the head, so a blend of the two pages puts volume at the cheekbones, which is the one placement wrong on both. Read that section and keep the part that is safe whichever row you turn out to be.

The calculator handles the same ground in the open. It scores all five readings at once, returns a margin against every row rather than a single word, and when the runner-up reaches 0.60 of the leading probability the result names both shapes — the middle widths above, entered with the field left blank, come back as exactly that: 0.3954 with the runner-up at 0.3824, a ratio of 0.967, printed as a blend rather than as a verdict. A result that names two shapes is not the tool failing to decide. On this pair it is the tool reporting where your numbers actually sit.

• Measure the four widths first and the angle never: forehead, cheekbones, jaw, length, straight across, hair off the forehead.
• Find the column in this page's first table matching your cheekbone width. If your forehead reading and your jaw reading both fall inside it, the tape has said what it can say.
• Feel for the corner below the ear and look at the lower edge of the jaw: flat and angular, or continuous and curved. No number involved.
• If you use the calculator, leave the angle field blank rather than guessing at it, and read the marker on the result that says which value was estimated for you.
• If the result names both shapes, style and buy for the half that is safe on either side, and read the borderline section on the round hairstyle page instead of merging the two pages.

The tape positions, the order to take them in, and what to do when a number is refused: How to measure your face, step by step ↓

What This Page Cannot Prove

The two rows are not two measured populations. In the fifty photographs this site calibrated its table on, the round row and the square row were each anchored by measured group means, and the site's own published note describes the groups behind those two rows as five or more labelled examples each, while the table it prints for the round row says ten faces and the row for the square says six. That is the entire evidence base for the centres on this page: sixteen faces out of fifty, labelled by one person, with no second independent pass and no published agreement figure. Nobody has checked whether a professional stylist given the same fifty would put the same labels on them, and this site publishes no accuracy number derived from that set because the set cannot support one.

The tolerances are one step removed from those faces. The spread on every row is the pooled within-group spread of the four groups in that same fifty with five or more members, and the resulting single figure travels across all seven rows, including these two. That is a reasonable way to build a band and it is not a measurement of how much round faces vary, or square faces vary, in the world.

The corridor in the first table is a property of chosen band widths, and this page says so twice rather than once. One tolerance for the angle row is 7.66°, a value carried across from the other rows; widen both bands to two tolerances and the empty strip closes into an overlap of 6.6°, and the claim that the two rows are disjoint at the angle goes with it. What survives either band width is that this row carries the pair's widest gap on both scales, and that no external page in six years has put a number on the same row.

The 2021 study is not evidence about face shapes and is not used as if it were. It measures bone on a radiograph in people sorted by growth pattern; the outline a mirror shows is soft tissue over that bone, which is a different quantity wearing a similar name. Its use here is to show that the number a styling table puts on one of these two faces sits below every group mean that study reported, which is a question to ask of the table rather than a verdict on any face.

And the step nobody has taken, here or anywhere in the table above, is the perceptual one. This page says where your tape reading falls against a published table. It does not say that a stranger can see the difference between these two labels, or that an 8.68° strip is visible rather than administrative. There is no study, ours or anyone else's, that maps these ratios onto what other people actually judge, so any sentence on this page about what you look like would be a guess wearing a number. What would change this page: labelled photographs at five or more per row for every row, a second rater over the existing fifty, or a perceptual result showing that one of these five checks is visible to other people. Until then the usable part is the shared ground, the blank field and the two rows' own limits, all of which are printed above rather than hidden below it.

Questions this page answers