glasses frame size & fit calculator

how the numbers are derived.

the short version, then the full method underneath for anyone who wants to check our working.

the short version

a photo has no ruler in it. to turn what the camera sees into millimetres, we need something in the picture whose real size we already know. we use your iris - the coloured ring around your pupil - because it is always there, it is round, and it does not change with your age. we take it as 13.2mm across - the figure that makes the tracker's reading agree with a face measured by ruler.

that is an assumption, and it is the weak point. irises vary by about half a millimetre between people. that sounds like nothing, but everything else is scaled from it, so half a millimetre of iris is about six millimetres of face. taking more readings does not help with that one, because your iris is the size it is and every reading is scaled by the same wrong number. the bank card is what fixes it.

a second error sits on top, and unlike the first it moves between captures - which is why one face can read 64mm one minute and 58mm the next. three things subtract from a reading and none of them add to it: an iris landing on too few pixels reads wider than it really is, which makes every width smaller; a head turned away from the camera shortens the distance between the pupils; and looking at anything close converges your eyes, which genuinely narrows your pupillary distance against the figure on a prescription. so repeat readings of one face sit at or below the true number rather than scattering either side of it. face the camera square, and look at the lens rather than at your own image.

a bank card fixes it. every card is exactly 85.60mm wide, by international standard. rest one flat against the side of your head, level with your eye, so it sits in the same plane as your pupils. mark its edges and we can measure your actual iris instead of assuming it. that removes the difference between your iris and the average one, which is worth a few percent on every width - though the card reading itself repeats less tightly than the assumed one does, so it corrects a bias rather than sharpening a number

where you hold it is most of whether it works. the card is measured against your iris in the same photograph, so anything that changes the card's size in pixels without changing your iris goes straight into the answer. holding it against your cheek does exactly that: the cheek curves forward, which puts the card perhaps 30mm nearer the lens, and a card nearer the lens is a wider card. measured on one face over three sessions, the cheek hold read 11.3mm and 11.5mm where the same face held against the side of the head read 12.1, 12.3 and 12.3. that is 8%, and it is larger than every error the corrections are there to remove. the white dashed line on the screen is drawn half a card height below your pupils for exactly this reason: rest the card's bottom edge on it and its middle lands in the pupil plane, without anyone having to estimate a middle. it is the single best thing you can do for accuracy here.

we correct for the shape of a head. the face tracker marks points on your outline as the camera sees it, and those sit slightly inside where an optician would put a ruler, because a head is round and a photo is flat. each width carries a small correction for that. those corrections were worked out against one person's measurements, so treat them as good to a few millimetres rather than exact.

the fit window comes from your temple track - the width of your head where the arms rest. a frame front sits inboard of that and the arms flex outward to reach your ears, so the range we look for is your temple track minus 4mm to minus 20mm.

what we will not tell you. where the lens centres should sit vertically, or how far down the lens your pupil lands. those depend on how a particular frame sits on your particular nose, and no photograph taken from the front can know that. an optician measures it with the real frame on your face. this is sizing, not dispensing.

the full method

every constant below is quoted from the code that runs above, and every assumption is named as one.

the scale: why 13.2mm

a photograph has no units. before a pixel can become a millimetre, something of known size has to be in the picture. the iris is the only thing that qualifies: it is always there, it is round, and it does not change size with your age. we take it as 13.2mm across and scale everything else from that:

millimetres per pixel = 13.2 ÷ iris width in pixels

13.2mm is not the width of a human iris, and it is worth being plain about that. a real adult iris is nearer 11.8. what the face tracker returns is the width of a circle it fits to yours, and that circle sits a little outside the coloured ring. 13.2 is the number that makes the reading agree with a face measured by ruler, which is the only thing the scale has to do. it was fitted against a known pupillary distance, two bank card calibrations and a face measured with a card in shot, and it lands within a millimetre on face width.

the catch, and it is the same catch: what the tracker fits is not fixed the way an iris is. it narrows when you smile, because your eyelids close in on it, and every width on the page moves with it. that is why the capture will not run while you are smiling. and adult irises do genuinely vary by about half a millimetre either way, which is a steady bias rather than random noise, so taking more readings cannot average it out. only the card removes it.

a second error sits on top of that one and does move between captures, which is why the same face can read 64mm one minute and 58mm the next. three things subtract from a reading and none of them add to it. an iris landing on too few pixels reads wider than it is, because blur spreads its boundary outward, and pupillary distance is inversely proportional to iris width - so a small iris shrinks every figure. a head turned away from the camera shortens the pupil separation by the cosine of the angle while barely touching the iris. and looking at anything close converges the eyes, which genuinely reduces pupillary distance by three to four millimetres relative to the distance figure an optometrist writes down. the gates below refuse the first two. the third cannot be gated, because no photograph can tell where somebody is looking - which is why the camera asks you to look at the lens rather than at your own face, and why a reading is better trusted at its highest than its average.

that is why the bank card option exists. every bank card in the world is 85.60mm wide - it is fixed by a standard, iso 7810. mark its two edges and the assumption is replaced by a measurement of your own eye:

your iris = iris width in pixels × (85.60 ÷ card width in pixels)

the calculator accepts a result between 10.2 and 15.1mm and refuses anything outside, on the grounds that a figure beyond that means the marks are not on the card. the band is wide, and deliberately not a statement about human anatomy: what the tracker returns is the width of a circle it fits to your iris, which is close to the real thing but not the same, and a card held against the side of the head photographs slightly small - so the figure runs a little high. two readings that agree are what makes a calibration trustworthy here, not the band.

the correction factors: 1.00, 1.00, 1.10

the face tracker returns points on your outline as the camera sees it. those are not always the points a dispenser would put a ruler on, so each width carries a correction - though two of the three are now 1.00, meaning no correction at all:

face width and temple track were 1.07 until they were checked against real frames. the 1.07 came from a ruler laid across one head above the ears, which reads the widest part of the skull rather than where an arm actually rests. three frames measured while worn, with a bank card in shot for scale, put that head at about 144mm rather than the 155mm the ruler claimed - and the frames say so without any photograph, because two of them sit 12mm inside the head and have to spring open to go on while the third is already 2mm wider and grips nothing. at 155 all three would grip. so the raw camera spans were right and the correction was the error. both are now 1.00. this is still one head, and a second would tell us whether that generalises.

face width = pixel width × scale × 1.00
temple track = pixel width × scale × 1.00
bridge = pixel width × scale × 1.10

the three differ because the mesh sits differently against each one. across the nose it reads high and the correction is doing two jobs at once - moving from the mesh's span to the nose, then from the nose to the frame bridge that clears it, which is always wider. across the face and the temples it reads low, because the face-oval contour sits inside the true head silhouette - so those two carry the same correction upward.

these were fitted to one person, against physical objects. temple track against three frames measured while worn, with a bank card in shot for scale, which put that head at about 144mm. the ruler reading of 155mm that the earlier figure rested on is the one the frames falsified, so it is not evidence for anything here. bridge against a frame's 21mm bridge, confirmed by a second frame at 22mm. face width is the same span again: the field feeds the temple track when that is left on “same as face”, so both are temple-to-temple measurements and both carry the same factor. it is not a cheekbone measurement, whatever the name suggests. an entered pupillary distance overrides the scale they all depend on, because a figure off a prescription beats any factor.

depth correction and the 400mm reference

move closer to the camera and the effect above gets stronger, so the same face measures narrower. we used to correct for that by scaling every reading back to a 400mm reference, using the distance the face tracker reports.

that correction is currently switched off. the number the tracker gives is not a distance. it solves the head pose against an assumed camera and its own canonical head, so it carries the ratio between the assumed lens and the real one. on laptop webcams it lands close to the truth; on a phone front camera, which is much wider, it reads roughly 1.6 times high - 52cm at a real 32cm. no browser exposes a focal length or a field of view, so it cannot be put right.

applying a correction with a wrong distance is worse than applying none, so we apply none. that is the end of the transformation matrix as a distance source, and it stays switched off. what replaced it does not ask the tracker anything, and is described below. an entered pupillary distance sits above both of them, because it re-scales everything to a figure that does not depend on distance at all.

the distance gate, and the correction that follows it

the iris sets the scale at the pupil plane. your cheekbones and temples sit behind that plane and your nose sits in front of it, so holding the camera closer strengthens the perspective and the same face measures narrower across the temples - while the pupillary distance, which is in that plane, does not move at all. over 28 captures on two devices, face width swung 21.6% with arm extension while the pupillary distance held steady. that is the largest single source of the spread between sessions.

the browser will not say how far away you are, but it does not have to. the iris is a fixed physical size, so how many pixels wide it is as a fraction of the frame is a direct distance signal:

distance fraction = iris width in pixels ÷ video width in pixels

two things are built on that. a gate refuses any capture above 0.028, which is roughly half an arm's length, because past that the reading is wrong however square your head is. and a correction normalises face width and temple track back to a reference distance of 0.0215:

gain = (1 + 6.13 × fraction) ÷ (1 + 6.13 × 0.0215)

it is applied to face width and temple track only. not to the pupillary distance, which sits in the plane the scale is read from and does not foreshorten; and not to the bridge, whose landmarks sit forward of that plane and move the other way, so correcting it would push the figure further from the truth rather than nearer it.

6.13 is one number standing in for two. fitted per device it comes out at 6.64 on a phone and 3.90 on a laptop, and the difference between them is the focal length no browser exposes. the pooled value is used because the wrong-device constant never makes a device worse than no correction at all - out of sample, the phone went from 3.2% error to 2.0% and the laptop from 4.0% to 2.6%. your own distance fraction is printed on the quality line as dist, so you can see what your captures are running at.

how the pupils are found

the face is tracked with google's mediapipe face landmarker, which returns 478 points on every frame. the first 468 map the face. the last ten come from a separate pass that finds the irises specifically, and those ten are the ones that matter here. 468 is the right iris centre and 469-472 its boundary; 473 is the left centre and 474-477 its boundary. pupillary distance is measured centre to centre. iris width is twice the average distance from a centre out to its four edge points. averaging four points rather than trusting one is what keeps the ruler steady from frame to frame.

pd = distance(468, 473) × scale
iris width = 2 × mean distance from centre to its four boundary points

the widths use named mesh points rather than the silhouette: 234 and 454 for the cheekbones, 127 and 356 for the temples, 245 and 465 for the nose at the bridge.

why the crosshairs never move

this one is worth stating plainly, because it looks like the opposite of what happens. the two crosshairs are drawn at fixed positions - one either side of centre, separated by exactly the pupillary distance at the diagram's millimetre scale. they are not tracking your eyes. the photograph is moved to them.

on capture, the detected pupils give a separation in image pixels and an angle. the photo is then scaled so that separation equals the pupillary distance in diagram units, rotated so the pupil line is level, and translated so the pupil midpoint lands on the crosshair midpoint.

photo scale = (pd × diagram scale) ÷ pupil separation in pixels
photo rotation = −atan2(Δy, Δx) of the pupil line

that ordering is deliberate. the diagram is drawn in millimetres and is the reference; the photograph is the thing fitted to it. it is why a frame drawn over your face is at true scale rather than approximately sized, and why dragging the photo afterwards cannot corrupt the measurements - it moves the picture, not the numbers.

the fit window, and the three widths

frame width is measured across the outer edges of the endpieces. it is not printed on the frame. two different numbers come off your temple track, and they answer different questions.

searched = temple track − 20mm to temple track − 1mm

that is the range we look through. it is deliberately wide, so a frame a little outside your ideal still appears, marked as a near miss rather than hidden.

snug = track − 10mm  ·  balanced = track − 7mm  ·  proportional = track − 4mm

those three are what we actually recommend, and they sit in the upper part of the searched range rather than spread across all of it. they are set from frames measured with a ruler across the front and worn daily - a 141mm front and a 139mm front on a head measuring about 144mm across the temples. a front a few millimetres inside the width of your head is what reads as the right size; much narrower and it looks like a smaller frame rather than a closer-fitting one.

front width is about how a frame looks, not how it feels. the same two frames make the point: one is 141mm at the front with its arms sitting 137mm apart and feels snug, the other is 139mm at the front with its arms at 152mm and feels loose. two millimetres apart at the front, fifteen at the arms. the arms are what press, and an optician can adjust them in five minutes.

this is set from two frames on one head, which is honest rather than authoritative. if you own a pair that fits and looks right, enter its width - the calculator will tell you what temple track that implies, and you should trust your own pair over the camera.

eye size and the 18mm

eye size = (frame width − 18 − bridge) ÷ 2

frame width is two lenses, one bridge and two endpieces - the corners carrying the hinges. the 18mm is what the endpieces actually come to across this catalogue: the median of frame width minus (two lenses + bridge) over every frame whose front width we hold, and the commonest single value in it. it is measured from the frames rather than assumed from a material. thin metal runs tighter and heavy acetate wider, so treat the derived eye size as a guide to what to search for rather than a specification.

lens height and fitting height

the minimum of 34mm is a progressive rule in disguise. what a corridor needs is fitting height - pupil centre to the lowest point of the glazed lens - and the pupil sits roughly 55-60% up the lens on a typical fit:

fitting height ≈ 0.55 × lens height

so 34mm gives about 18mm of fitting height, which suits most modern progressives. designs with a shorter corridor - the channel the prescription changes through - will go lower.

the calculator does not offer an optical centre or a fitting height of its own. those depend on where the frame actually ends up sitting on your nose: bridge height, the profile of your nose, and pantoscopic tilt, which is how far the frame front leans back from vertical. a photograph taken from straight on cannot tell you any of them. that measurement belongs on the face with the real frame.

the capture protocol

a single frame is not a measurement. the reading is gated before it starts and taken across many frames:

head roll ≤ 8° · yaw asymmetry ≤ 0.06 · iris ≥ 20px · distance fraction ≤ 0.028 · face centred in frame
then a 3 second settle, then 24 frames, and the median of each measurement

the middle value rather than the average, so one blink or wobble cannot drag the result. how much the 24 frames disagreed is shown alongside as a quality figure. a tight spread means you held still - it does not mean the answer is right. the bank card step gates tighter at 3° of roll, because a tilted card puts its 54mm height into a horizontal measurement at about 1% per degree.

face shape

eight shapes, decided by comparing four distances: brow width, cheekbone width, jaw width and face length. if the face is more than 1.5 times longer than it is wide, it is long. if the jaw is within 3% of the cheekbones, it is square or rectangular depending on that same ratio. a brow wider than the cheekbones with a narrow jaw is heart or triangle. and so on. it is a starting point for shape preference, not a clinical classification, and it is overridable - as is the colour suggestion, which samples forehead and cheek for skin, the iris interior for eye colour, and the scalp or brows for hair, then sorts warm against cool and light against deep.

the 36 styles

every frame in the catalogue is filed under one of 36 styles, which group into nine base shapes. the split is not even - the panto family alone is nearly half the catalogue, because that is what independent makers actually make.

boston 9 styles · 121 frames

the panto family. a soft keyhole-ish shape, slightly wider than tall, flat along the brow and rounded below. the safest shape on most faces, which is why it dominates the catalogue.boston · boston - keyhole · boston - square · panto · panto - bold · panto - round · panto - slim · panto - square · panto - wire

wellington 3 styles · 36 frames

the wayfarer outline - a trapezoid, wider at the top than the bottom, with a strong brow line. reads bolder than a panto at the same width.wayfarer · wayfarer - squared · wellington

round 6 styles · 31 frames

a true circle or close to it. draws attention to the eye and softens an angular face; wire versions almost disappear, acetate versions do the opposite.round · round - curved bridge · round - keyhole · round - large · round - saddle bridge · round - wire

browline 2 styles · 21 frames

heavy across the top, light or rimless below - the weight sits on your brow and lifts the face. the shape everyone recognises from the 1950s.browline · browline - slim

square 5 styles · 20 frames

corners kept, edges straightened. adds structure to a round or soft face and reads more formal than a panto of the same size.square - bold · square - geometric · square - oversized · square - rounded · square - soft

cateye 2 styles · 17 frames

the outer top corner sweeps up. lifts the eye and the cheekbone; the sharp version does it emphatically, the soft version barely at all.cat-eye · cat-eye - sharp

aviator 5 styles · 12 frames

a teardrop that is wider at the top and drops below the eye line. built for metal and for sun, and the only family here that regularly sits outside a rectangle.aviator - flat top · aviator - metal · aviator - navigator · aviator - round · aviator - teardrop

rectangle 2 styles · 3 frames

longer than it is deep, with square corners. the shallowest family, so worth checking against the lens height minimum if you wear progressives.rectangle · rectangle - slim

oval 2 styles · 2 frames

a circle stretched sideways. quiet and hard to get wrong, and rarer in current production than you would expect.oval · oval - wire

the style decides which drawing appears on the card and on the diagram. the base shape is what the face-shape guidance works with, so two styles in the same family are treated interchangeably when we suggest what suits you - the difference between them is a matter of taste, not of fit.

what changes when you try a frame on

more than you might expect, and less than would be ideal. three things happen:

the drawing is swapped for that style's own outline
the two halves move apart by (bridge − 20mm) × scale
the whole thing is pre-shrunk so the total still equals the real frame width

that third step is the one worth understanding. widening the bridge would otherwise make the frame wider overall, so the artwork is scaled down by exactly enough to absorb the gap. the lenses genuinely get narrower as the bridge gets wider - which is what happens on a real frame at a fixed outer width, and it is why a 49□21 and a 51□19 can both measure 143mm across.

so the outer width is true, the bridge gap is true, and the lens widths follow correctly from both.

what is not true to the frame is the lens depth. the drawing keeps its own outline's proportions vertically rather than scaling to the lens height we hold for that model. the height figure is still used - it is quoted on every card, it feeds the 34mm minimum, and it is on the printed sheet - but the picture does not reflect it. a shallow rectangle and a deep panto of the same width will draw with the same depth.

we would rather say so than let the drawing imply a precision it does not have. it is on the list to fix: the ratio is already in the catalogue for the frames whose makers publish it, so the drawing could be made height-accurate for those and derived for the rest.

what this is not

it estimates sizing from a photograph and a set of assumptions, several of which are named above as fitted to one person. it is not a dispensing measurement, it produces no optical centre, and it cannot see a nose in profile. confirm dimensions with the maker and have lenses fitted by a qualified optician.