Methodology
Every number Facemap reports is measured, deterministic and reproducible. The same photograph always returns the same result, and each result traces back to a distance, an angle or a residual taken on the 478 point landmark mesh. The overall score goes one step further: it is calibrated against a published research set of faces that published reference ranges for attractiveness, so it predicts how average people would rate your attractiveness out of 100 from that photo. This page publishes the whole pipeline, including the formula behind each measurement and how the score is built.
Analysis happens in your browser. Your image is not uploaded to our server.
A face detector runs live in your browser at roughly three readings a second. Before anything is measured, the frame is checked: exactly one face, head turn and tilt within a few degrees of straight on, a neutral closed mouth, and a face large and sharp enough in the frame. Unsuitable photos (head turned or tipped, camera too close or too far, whole head not in frame, or no face found) are rejected and the person is asked to scan again, rather than guessing a score.
Google's MediaPipe Face Landmarker places 478 three-dimensional points on the captured frame, including both irises. The model file is served from this site and the inference runs on your device. No image is uploaded.
Coordinates arrive normalised to image width and height separately, so the horizontal axis is rescaled by the image aspect ratio before any distance is taken. The mesh is then rotated back towards a straight-on view to remove residual yaw, pitch and roll, and every distance is expressed relative to your own face, usually bizygomatic width, eye width or interpupillary distance, so camera distance and image size cancel out.
22 measurement categories are derived from the corrected mesh: distance ratios, angles in degrees, coefficients of variation and mirrored-pair residuals. Together they involve over 1,600 individual calculations. Every formula and every landmark index used is published below.
Each reading carries a confidence value. Expression, residual head angle and low landmark confidence in a region reduce the weight of the readings that depend on it, so a raised brow or a partly hidden forehead nudges the result far less than it otherwise would.
The camera automatically captures three photographs at different distances in a single scan. Each photo is measured independently, then your score is read from the clearest, most level photograph rather than a blend of all three, because each photo carries the pose of its own moment. This is what makes the score reproducible between scans.
Each individual measurement is reported against a published reference range, so you can see where that feature sits. The reference comparison uses a Gaussian match curve: a reading on the published reference reads highest, and the reading falls away smoothly as it moves outside the tolerance band. Individual readings are clamped between 1 and 99, because no single measurement should ever claim perfection.
score = round( 100 x exp( -0.5 x ( (reading - target) / tolerance )^2 ) ), clamped to 1 to 99
The headline score is not a simple average of those reference comparisons. The corrected, confidence-weighted readings are combined and mapped onto a 1 to 100 scale by a model calibrated against Facemap scans compared against published reference ranges face photographs, each rated for attractiveness by around 60 raters. The weights and the curve come from what those panels actually scored, so a high number means your measured structure resembles faces that people rated highly, and the percentile you are shown is your position in that Facemap scans.
To ensure the score is stable and reproducible, Facemap weights measurements based on how much they are affected by small changes in pose or expression.
Full Weight (Steady)
Measurements that stay steady between repeat photographs carry full weight: jaw angle, jawline, face shape, midface, nose proportions, eye spacing, chin, face width-to-height, and mouth-to-nose.
De-weighted (Moveable)
Readings that move between repeat photographs are reported in full but de-weighted so they cannot swing the headline score: lip fullness, facial thirds, eye-size symmetry, canthal tilt, eye-to-brow distance, lip ratio, facial fifths, overall symmetry, brow symmetry, and gonial angle.
Landmark numbers refer to MediaPipe Face Landmarker indices. Point 1 is the nose tip, 152 the chin, 234 and 454 the cheek edges, 468 and 473 the iris centres.
Eyes and brows
Angle of the inner-to-outer eye corner axis, averaged left and right.
mean( atan2( y(inner) - y(outer), |x(outer) - x(inner)| ) ) in degrees
Basis: Anthropometric norms for neutral to mildly positive canthal tilt.
Intercanthal width expressed in eye widths.
dist(133, 362) / mean( dist(33, 133), dist(263, 362) )
Basis: Published reference range for eye spacing.
Vertical brow clearance, normalised to eye width.
mean( |y(105) - y(159)|, |y(334) - y(386)| ) / eye width
Basis: Aesthetic surgery references for brow position over the orbital rim.
Pupil separation as a fraction of your own bizygomatic width.
dist(468, 473) / dist(234, 454)
Basis: Adult interpupillary to face width ratios in anthropometric data.
Difference in aperture openness between the two eyes.
|EAR(left) - EAR(right)| / mean EAR, where EAR = eye height / eye width
Basis: Openness is compared rather than raw area so a few degrees of head turn does not register as asymmetry.
Eye openness, palpebral height against width.
mean( dist(159, 145) / dist(33, 133), dist(386, 374) / dist(263, 362) )
Basis: Typical open-eye aperture ratio in front-facing photography.
Jaw and chin
Interior angle at the jaw corner.
angle at 172 between 234 and 152, mirrored at 397, then averaged
Basis: Cephalometric gonial angle ranges reported in the maxillofacial literature.
Bigonial width against bizygomatic width, the cheek-to-jaw taper.
dist(172, 397) / dist(234, 454)
Basis: Taper ratios associated with a defined but not square lower face.
Chin height against philtrum length.
|y(152) - y(17)| / |y(0) - y(2)|
Basis: Lower-face vertical balance norms.
Nose, lips and midface
Pupil line to upper lip distance, in interpupillary units.
|y(pupil midpoint) - y(13)| / dist(468, 473)
Basis: Midface compactness ratio used in facial proportion research.
Nose width against nose height.
dist(129, 358) / dist(168, 2)
Basis: Classical anthropometric nasal index bands.
Mouth width against alar base width.
dist(61, 291) / dist(129, 358)
Basis: Published reference range for mouth-to-nose proportions.
Lower vermilion height against upper vermilion height.
|y(17) - y(14)| / |y(13) - y(0)|
Basis: Commonly cited 1 to 1.6 upper-to-lower lip proportion.
Proportion and ratio
Evenness of the three vertical thirds.
standard deviation of the three thirds / their mean
Basis: Published reference range for vertical facial divisions.
Evenness of the five vertical columns across face width.
standard deviation of the five columns / their mean
Basis: Published reference range for horizontal facial divisions.
Facial width-to-height ratio.
dist(234, 454) / |y(13) - y(9)|
Basis: fWHR as defined in the behavioural science literature, bizygomatic width over upper-face height.
Composite of length-to-width and cheek-to-jaw taper.
0.6 x (face height / face width) + 0.9 x (bigonial / bizygomatic)
Basis: Weighted blend chosen so oval reads highest and both very long and very square outlines fall away.
Symmetry
Mean mirrored-pair residual across the whole face.
midline fitted from 24 mirrored midpoints, head turn removed by regressing the left-right offset difference on landmark depth, then mean hypot(lateral residual, vertical residual) / face width
Basis: Residual asymmetry after perspective correction; no face measures zero.
Mirrored-pair residual across the brow region only.
same residual as facial symmetry, restricted to the brow pairs
Basis: Region-restricted version of the whole-face residual.
Mirrored-pair residual across the mouth region only.
same residual as facial symmetry, restricted to the lip pairs
Basis: Region-restricted version of the whole-face residual.
Combined vermilion height against mouth width.
|y(17) - y(0)| / dist(61, 291)
Basis: Vermilion height to mouth width proportions in aesthetic reference data.
All 21 measurements are computed from a single captured frame in under a second, on your own device.
The free scan places the mesh, computes all 22 readings and returns your Attractiveness Score. No account, no card.