Why Your Canthal Tilt Reading Changes When You Tilt Your Head: What Official Portrait Photos Show

Upload two selfies of the same face to a canthal tilt tool and you can get two different verdicts. One reads neutral, the other slightly negative. The face is the same in both. For this measurement, the difference between two photos is big enough to flip the verdict. The reason is scale. Canthal tilt is the angle of the line running from the inner corner of the eye to the outer corner, measured against horizontal. In the official portrait photographs I measured for this write-up, the readings ran from −0.1° to 10.1°, a ten-degree window. The median was 4.4°, and the middle half of the faces fell between 3.3° and 5.3°. The gap people argue about, slightly positive versus slightly negative, is two or three degrees. A head that is tilted or dropped by a few degrees moves the reading by the same amount as the thing you are trying to measure. How the angle is computed A face landmark model places points on the eye corners. MediaPipe's Face Landmarker is one common option, and its face mesh model outputs 478 three-dimensional landmarks. Pick the inner and outer corner of each eye and the tilt is a single arctangent. A simplified version (landmarks arrive normalized to the image, so convert them to pixels first, or a 3:4 phone photo stretches the angle): // Pixel coordinates: y grows downward, so inner.y - outer.y > 0 // means the outer corner sits higher than the inner corner. function eyeTiltDeg(inner, outer) { const run = Math.abs(outer.x - inner.x); return (Math.atan2(inner.y - outer.y, run) * 180) / Math.PI; } const left = eyeTiltDeg(leftInner, leftOuter); const right = eyeTiltDeg(rightInner, rightOuter); const tilt = (left + right) / 2; // the number people quote const gap = left - right; // left-right difference Notice what the line is measured against: the edge of the photo. If the head is rolled, the face's own horizontal and the photo's horizontal disagree, and that disagreement goes straight into each eye's reading. Illustration of an adult woman tilting her head toward a phone camera. Head roll: 2° of error per degree, cancelling in the average Tilt your head sideways and one outer eye corner goes up while the other goes down. Across those portraits, each degree of roll changed the left-right difference by 2.0° (±0.1°), almost exactly what geometry predicts. The average of both eyes barely moved, at 0.03° per degree, because the two shifts are equal and opposite. One portrait shows it clearly. The head was tilted 2.8°. Measured against the photo edge, the two eyes read +8.1° and +2.7°, a 5.4° difference that is almost entirely the tilt. Two practical consequences follow. First, quote the average of both eyes, not either eye alone. Second, if you want to say anything about left-right asymmetry, you have to level the face before measuring. Leveling against the eye corners themselves would be circular, so use other point pairs. We used 12 matching pairs that do not include the eye corners. Here is what that did to the left-right difference, grouped by how much roll each portrait had: Roll in the photo Difference before leveling Difference after leveling under 2° 1.9° 1.2° 2° to 4° 6.5° 1.1° over 4° 11.4° 1.6° After leveling, the difference stopped following the roll (0.00° ± 0.08° per degree). Head pitch and turn Pitch is harder. Portraits with the chin lower read lower, by about 0.25° per degree of pitch (±0.1°). Treat that as a rough estimate: it comes from comparing different faces photographed at slightly different pitches, not from one face measured at several angles. A ten-degree nod is easy to do without noticing. It is worth about 2.5°, which is more than half the width of the commonly quoted male range of 0° to 4°. You cannot undo pitch the way you undo roll, because a single flat photo does not carry a reliable record of the true angle. The honest options are to estimate how much it is likely pulling the reading and say so, or to refuse photos that are too far off. Within the roughly ±11° of left-right turn present in the portraits, turning the head produced no measurable shift in the average (0.03° ± 0.13° per degree). Beyond about 15° a reading deserves low confidence, and past about 25° it is better refused than turned into a number nobody should trust. The noise that is left After leveling, a gap of about a degree between the two eyes remains. It is tempting to read that as real asymmetry. The data argue against it: in almost every portrait the subject's left eye read slightly higher, by 1.2° on average. Real asymmetry would not lean the same direction that consistently, so part of that leftover is the tool. Repeatability gives a second floor. Rotating the same photo by ±3° and re-running the analysis moved the reading by about 0.5° typically and by up to 1.4° in the worst case. That wobble comes from the detector landing a pixel or two differently each time. Put together, one photo resolves canthal tilt to roughly ±1°. A result within about a degree of zero is a neutral result, and calling it positive or negative is reading noise. What this means when you measure yourself Illustration of a phone on a tripod at eye level. Shoot at eye level with the phone held straight and look at the camera. Compare only photos taken the same way, because half a degree between two different setups says nothing. Quote the average of both eyes and ignore left-right differences under about a degree. And be suspicious of any tool that gives you a number to a tenth of a degree without saying what it did about head pose. I built the measurement described here into LooksmaxScale's canthal tilt tool. It levels roll before measuring, reports the roll it found, estimates how much pitch is likely pulling the reading, and declines photos beyond the limits above. The measurements in this article come from the same code. A set of portraits is not a population, and the pitch figure in particular is a correlation across different people. The roll result is the firm one, because it matches the geometry almost exactly.