How much do face measurements move? A reproducible study on a public face set

Every face-analysis tool shows you a number. Almost none shows you how much that number moves when nothing about your face has changed. This study measures that. We took 102 adults from a published research set, ran them through the same 478-landmark engine that powers our free tools, changed the photo in six ways, and recorded what shifted. We also tested whether the score agrees with 2,513 people's ratings of the same faces. Some results are reassuring. Others are the reason we print a warning next to our own score.

The common format is a company counting the photos its own users uploaded: how many ovals, what the average score was. We cannot publish that, and we would not want to. The measurement tools run in your browser, the photo is not uploaded for the measurement, and the analytics events we keep carry no scores (see privacy). A dataset of users' faces does not exist on our side by design.

There is a second reason. A sample made of people who chose to have their face scored is not a sample of faces: it is a sample of people worried about faces. Its average says little about anyone else, and nobody can re-run it, because the photos are deleted. So we used a set that anyone can download, and we kept the question narrow enough for it to answer.

The faces are the Face Research Lab London Set (DeBruine & Jones, 2017): 102 adults photographed straight on under standard studio conditions, released under a CC BY 4.0 licence, together with 1-7 attractiveness ratings from 2,513 raters, each of whom rated every face. We used the neutral front-facing photo of each face as the reference, plus the same people's smiling photo and left three-quarter photo from the same release.

Landmarks come from the same MediaPipe release and model file the live site loads, so the numbers here are the numbers a visitor would see. Because we run exactly the code of the tools, a change to the engine changes these results; that is why every table below is generated by a script rather than typed.

Search volume for "average face" is real, so here are the numbers, kept short because the full tables live on the measurement statistics page. Each cell shows the median and, in brackets, the range containing the middle 80% of the set.

Two cautions. First, "average" is a statistical property of this set, not a goal: averageness does correlate with how attractive faces are rated in some studies (Langlois & Roggman 1990; Rhodes 2006), but that is a finding about composites and populations, not an instruction to move your measurements towards a median. Second, the set is 102 young adults, 69 of them white, photographed in one place: it should not stand in for a population.

For each condition we measured the same 102 faces again and took the absolute difference from the neutral photo, face by face. The table reports the median difference and the 90th percentile (the size of the shift that one face in ten exceeded), so a reader can see both the typical and the bad case. Four conditions are simulated by editing the neutral photo; two are real photographs from the set.

Re-saving a photo smaller or darker

Shrinking a photo to 338 pixels wide or darkening it to under half its brightness moved the 1-10 score by a median of 0.3 and 0.4 points, with one face in ten shifting by about 1.2 to 1.0. That is a small shift but not a zero shift, and it comes from the landmark model placing its points slightly differently on a different file. It also shows how little a score moves when only the file changes, which is the lower bound on what you should read into any difference.

A tilted photo that should change nothing

Rotating a photo in its own plane does not change any distance on a face, so in principle a 5° tilt should change no ratio and no label. In practice the face-shape label changed for 17 of 102 faces at 5° and 23 at 10°, and the eye-shape label for 11. Those changes are the landmark model's own noise crossing a threshold. It is the single most useful number for anyone who has been told "you are an oval": a classification boundary is a line drawn through a continuous measurement, and a face near the line flips when the photo barely changes.

Smiling

A smile is the largest change that is not a deliberate pose. The 1-10 score moved by a median of 1.3 points and by more than 4.5 for one face in ten; the face-shape label changed for 41 faces and the lip-shape label for 54, which is what you would expect, since a smile changes the mouth and the lower face. Pupillary distance moved by a median of 0.9 mm and canthal tilt by 0.6°. If you measure yourself, measure with the same expression every time.

Turning the head

The three-quarter photos are a deliberate stress test, not a fair use of a front-facing tool. Symmetry moved by a median of 85.5 points on its 0-100 scale, the 1-10 score by 4.5, and the face-shape label changed for 92 of 102 faces. The tool's readings simply stop describing the face. The live camera flow checks distance, centring, head turn and tilt before it captures a photo, for exactly this reason (the check is part of the camera component, not of the measurement itself).

Running a published set through the engine did more than characterise it: it found faults that our own test photos had hidden. Each one changed what the live tools report, and each is in the changelog.

• An aspect-ratio error. The landmark model reports horizontal positions as a fraction of width and vertical positions as a fraction of height, and the engine had been treating both as the same unit. Every photo that was not square was measured on a distorted face. Cropping the same 102 faces to 3:4 changed the face-shape label of 27 of them and made 59 come out round. The fix converts both axes to one unit before any distance is taken. Our test photos were square, which is why nobody saw it.
• Thresholds that did not generalise. The first face-shape thresholds came from 32 faces, most of them synthetic. On real faces no one came out heart-shaped or oblong, and 30 of the 53 men were labelled triangle. The thresholds were rebuilt from the percentiles of this set.
• A mouth label that fired for almost everyone. A neutral mouth measures a median of −3° at the corners, and the cut-off for "downturned" was 0°, so 82 of 102 faces were labelled downturned. The cut-off moved to −6.5° and the count fell to 19.
• An iris that reads large. The model's iris diameter ran a median 5.9% above the hand-marked limbus, so pupillary distance came out too small (median 58.4 mm). A ×1.059 correction brought it to 61.9 mm.
• Eye-spacing and nose cut-offs. On the same set 89 of 102 faces were labelled wide-set and none close-set, and none had a long nose. Those cut-offs were recalibrated from the set's percentiles.

Two limits on this. Several cut-offs (face shape, eye spacing, long nose) are percentile choices made on this set, so the distributions are in-sample: they describe how the detector divides these faces, not how common each trait is. And there is no labelled ground truth for face shape, so we claim no accuracy rate. Full thresholds and their sources are on the methodology page.

This is the question a face-analysis score implicitly answers, and almost no tool tests it. We tested ours against the 2,513 raters' scores for the 102 faces, using Spearman rank correlation with bootstrap 95% intervals.

The composite is 40% symmetry plus three proportion fits at 20% each, and on these faces it did not track the raters. An earlier result that seemed to show jaw width predicting ratings (ρ about −0.40) turned out to be sex mixed into the data: the women's faces were both rated higher and narrower-jawed, and within each sex single ratios were close to zero. The only measure with a small, positive signal was how near a face is to the average of its own sex, in line with the averageness findings (Langlois & Roggman 1990; Rhodes 2006), but it needs the person's sex, which the tool does not ask, and it faded with a smile.

We kept the formula, converted the score to a rank among these 102 faces so that a 5.5 means the middle of the set, and published the result on the tool pages. Read the score as a measurement of this photo against a research set, not as a verdict about a face. For the limits of the sample behind that sentence, and how forum rating tiers relate to it, see the looksmaxxing chart guide.

1. 1Compare a tool with itself, under the same conditions. Same distance, same light, same expression, same head position. The shifts above are the size of the noise you add by changing any of them.
2. 2Treat differences under about one point on a 1-10 score as noise. A re-saved or darkened copy of the same photo moved it by 0.3 to 0.4 points at the median and up to about 1.0 for the worst tenth.
3. 3Take two or three photos before you believe a label. A face near a classification boundary changes label on a barely different photo, so one "oval" or "downturned" is a single draw.
4. 4Do not compare your number across apps. Different models, different measurements, different reference groups. A score is a position within a reference set.
5. 5Do not read a score as an opinion about you. On this set, it did not predict what people said about the faces. It is geometry, and it is useful when you are checking whether something you changed, like a haircut or posture, moved a ratio you can see.

• ✓Small sample. 102 faces. Percentiles are broad, and a correlation could hide a relationship of about 0.2 in either direction. We do not publish 5th or 95th percentiles.
• ✓One set, one place. Studio photographs of young adults, 69 of 102 white. It says nothing about older faces, other photographic styles, or populations the set does not include.
• ✓Four conditions are simulated. A tilt, a shrink and a darkening applied to the neutral photo, so they test the landmark model's noise, not a real camera. The smiling and three-quarter photos are real.
• ✓What a front photo cannot measure. Profile angles, eye depth and the top third of the face are not measured from a front view, and the tools say so on each page.
• ✓Not a population norm. The cut-offs for face shape, eye spacing and a long nose were picked from this set, so counts are descriptive of the detector.
• ✓Per-face values are not published. The licence would allow it, but a table of identifiable people's measurements next to attractiveness ratings is not something we want on the web. The aggregates are on the measurement statistics page.

To reproduce the tables, download the neutral front, smiling front and left three-quarter archives and the info file from the figshare record (doi:10.6084/m9.figshare.5047666.v5), unzip them under /tmp/flondon, and run two scripts from the site repository: gen-photo-drift.mjs for the conditions table, and gen-measure-stats.mjs for the distributions. They compile the engine's source, extract landmarks with the same MediaPipe release and model the site loads, and write only aggregates. If the engine changes, both must be rerun, which is the reason the tables are generated, not typed.

“How to cite: The Face Analysis (2026). How much do face measurements move? A reproducible study on a public face set. Measured on the Face Research Lab London Set (DeBruine & Jones, 2017, CC BY 4.0), engine of 16 October 2026.