How the score is calculated

Landmarks and the scan

Everything starts from Google’s MediaPipe Face Landmarker, an open model that runs in your browser and places 478 points on a face in every camera frame: 468 points of a dense surface mesh and 10 around the irises. Each point has a horizontal and vertical position and an estimated depth.

The scan captures five poses: front, turned left, turned right, tilted up, tilted down. A pose is captured only when your head holds inside a target window, for example 0° of yaw within ±6° for the front view, for 0.6 seconds. Head pose comes from the facial transformation matrix the model returns with each frame; positive yaw means you have turned to your right.

Each capture is a burst of 8 consecutive frames. The frames are fused by taking the median position of every landmark, which removes single-frame jitter, and the spread across frames is kept so the report can show it next to each number: a canthal tilt of +4.1° ±0.4° means the eight frames agreed to within 0.4°.

All measurements are ratios or angles, so they do not depend on how far you sit from the camera. Lengths are divided by the span between the outer eye corners, the most stable width the model finds.

Six components

The score is built from six components, each scored from 0 to 100. The reference is a typical real face, which scores 50 on every component and 5.0 overall. “Typical” is defined in the formula: a left–right mismatch of 3% of the span between the outer eye corners, an overall shape 4% away from the average face, and every individual feature about 0.8 standard deviations from the population mean, which is the average distance for a normally distributed trait.

The profile needs the turned angles of a camera scan. With a single uploaded photo it is left out and the other five weights are rescaled to add up to 100%. Any profile measure the model can’t see clearly (confidence below 0.5) is dropped rather than guessed.

Symmetry and averageness reward a face for being closer to the population’s centre than a typical face is. This is the averageness effect: a composite made by averaging many photographs is rated more attractive than most of the faces that went into it (Langlois & Roggman 1990; Rhodes 2006), and making a face more symmetric raises its ratings too. Each is scored as

v = 50 + 50 · tanh((typical − measured) / s)

where “measured” is your left–right mismatch (symmetry) or your shape’s distance from the average face (averageness), “typical” is the typical face’s value, 3% and 4%, and s sets how quickly the score moves. A typical face scores 50. A perfect mirror, or a shape exactly at the average, scores 74, the most these two components can give. More mismatch than typical scores below 50.

Proportions, eyes, jaw and chin, and profile are each a weighted mean of sub-measures. A sub-measure is scored by how much closer to its ideal value it sits than a typical face does:

sub = 50 + 50 · tanh((T − D) / s) D = your distance from the ideal T = a typical face’s distance from it

As far from the ideal as a typical face scores 50; closer scores more, further scores less. The ideals are where the formula peaks, not targets for a face. Some come straight from the research: equal thirds, even fifths, and an eye spacing at the population average, which is where Pallett and colleagues found the optimum. Others sit deliberately beyond the range of real faces, such as a canthal tilt of 14°, so that across the range people actually have, a little more tilt, openness or eye width keeps scoring a little better. None of them describes a face anyone should aim for.

From components to a score

The weighted components give a composite C between 0 and 100. A logistic curve maps it to the 1.0–10.0 scale:

C = Σ wₖ · vₖ (weights sum to 1) score = 1 + 9 / (1 + e^−((C − 52.68) / 12))

The midpoint, 52.68 = 50 + 12 · ln(5/4), is chosen so that C = 50, a typical face, scores exactly 5.0. The curve is steepest in the middle and flattens towards the ends, so the step from 5 to 6 takes a smaller change in the composite than the step from 8 to 9. The score is rounded to one decimal.

The tiers are fixed bands of the score:

• below 4.0: Below average
• 4.0–5.4: Average
• 5.5–6.9: Above average
• 7.0–8.4: Well above average
• 8.5 and above: Exceptional

Four reference points:

• A typical real face scores 5.0.
• The population-mean face, MediaPipe’s canonical face, scores about 7.4 with no preference and 7.3 rated as a man or a woman. It is perfectly symmetric and exactly average, so both of those components give their 74, but on the other four it sits at the mean rather than at the peaks.
• The range the formula can produce runs from about 1.2 to about 9.1 with no preference and 9.3 as a man or a woman. The scale is labelled to 10.0, but no face reaches it, because symmetry and averageness stop at 74.
• Repeat scans of the same face, in our tests with simulated camera noise, moved the score by at most ±0.2.

Rating preference. Before the scan you choose to be rated as a man, as a woman, or with no preference. The choice changes two things, only where the research supports a difference between the sexes: the weights of eyes and jaw (Table 1), and the peaks of the sexually dimorphic sub-measures, such as jaw width and angle, chin width, brow height and eye size. With no preference, each of those peaks sits midway between the two.

Strengths and detractors. The report names up to three components scoring 52 or more as strengths and up to three scoring 48 or less as detractors, each with one specific sentence: what was measured, the value, and which way it pushes the score.

The research behind it

Attractiveness research does not agree on everything, but a few findings repeat across decades, cultures and methods. The score uses those and leaves the rest out.

• Langlois, J. H., et al. (2000). Maxims or myths of beauty? A meta-analytic and theoretical review. Psychological Bulletin, 126(3), 390–423. Raters agree strongly about which faces are attractive, within and across cultures, so there is a shared signal to measure.
• Langlois, J. H., & Roggman, L. A. (1990). Attractive faces are only average. Psychological Science, 1(2), 115–121. Composites of many faces are rated more attractive than most of the faces they are built from: the averageness effect.
• Rhodes, G. (2006). The evolutionary psychology of facial beauty. Annual Review of Psychology, 57, 199–226. Averageness, symmetry and sexual dimorphism as the three best-supported geometric correlates.
• Little, A. C., Jones, B. C., & DeBruine, L. M. (2011). Facial attractiveness: evolutionary based research. Philosophical Transactions of the Royal Society B, 366(1571), 1638–1659. A review of the same effects, with their sizes and their caveats.
• Perrett, D. I., May, K. A., & Yoshikawa, S. (1994). Facial shape and judgements of female attractiveness. Nature, 368, 239–242. The most attractive faces are not exactly average, which is why four components reward moving towards an ideal rather than towards the mean.
• Perrett, D. I., et al. (1998). Effects of sexual dimorphism on facial attractiveness. Nature, 394, 884–887. Preferences for masculine male faces are weaker and less consistent than often assumed, which is why the sex-specific weight changes are small.
• Cunningham, M. R. (1986). Measuring the physical in physical attractiveness. Journal of Personality and Social Psychology, 50(5), 925–935. Specific measurements, such as eye size and chin width, correlate with ratings.
• Pallett, P. M., Link, S., & Lee, K. (2010). New “golden” ratios for facial beauty. Vision Research, 50(2), 149–154. The most attractive eye spacing and eye-to-mouth distance are close to the population average.
• Kartynnik, Y., et al. (2019). Real-time facial surface geometry from monocular video on mobile GPUs. arXiv:1907.06724, and Google’s MediaPipe Face Landmarker documentation and model card. The landmark model, its accuracy and its known failure cases.

Limits

• Geometry only. Skin, hair, expression, grooming, body fat, age and everything that is not the shape of the face are invisible to it. People rating a real face see all of those.
• Not learned from ratings. The weights and ideals come from published findings, not from fitting a model to rated faces, and the score is not a percentile. The typical face that anchors 5.0 (3% mismatch, 4% shape residual, 0.8 standard deviations per feature) is a judgement centred on MediaPipe’s canonical face. It has not yet been re-centred on real scans, so the scale may shift when it is.
• One reference face. “Mean face” here means MediaPipe’s canonical face mesh. It is a single template, not an average built separately for every age, sex and ancestry, so averageness measures distance from that template.
• Landmarks are estimates. The model infers points it can’t see, such as a jawline under a beard or hair. It has few points under the nose, so the nasolabial angle reads wider than a clinical measurement and carries a low weight.
• The camera changes the face. A phone held close exaggerates the nose and narrows the face; light from one side shifts landmarks; a smile changes the jaw and eyes. The scan asks for a neutral face and checks distance and light, and the report prints the spread of every measurement.
• Not medical. Nothing here is a diagnosis or a recommendation for treatment or surgery.

If you want to see the formula at work on your own face, the attractiveness test shows the score and tier for free; the Full Face Report shows every component behind it.