How the Rate My Face score works

The photographs, measurements and scoring steps behind the Rate My Face score, its percentile and its tier

Updated
On this page

The Rate My Face score combines facial measurements from a frontal and a lateral photograph with an AI model's assessment of individual features and of the face as a whole.

Scoring pipeline

Rate My Face has two components: a geometry module and an AI model.

  • Geometry module. It detects facial landmarks on each photograph and calculates measurements from their coordinates with fixed formulas. The same landmark coordinates always give the same measurements.
  • AI model. A multimodal model that accepts images and text together. It receives the photographs, the measurements, sex and age, and returns the feature scores and the overall score.
  1. 01PhotographsFrontal and lateral
  2. 02Landmarks478 points per photograph
  3. 03MeasurementsRatios and angles
  4. 04AI modelPhotographs, measurements, sex and age
  5. 05Scores13 features and the overall score, 1–10
  6. 06Percentile and tierRank among faces of the same sex

The sections below follow the pipeline from left to right.

Input photographs

Overchat AI Rate My Face uses a frontal and a lateral photograph as input.

Why we use 2 images instead of 1

We use 2 images instead of one because the profile view adds features a frontal photo cannot measure, such as the angle between the nose and the forehead, chin projection and the convexity of the profile, allowing for a more accurate reading overall.

Facial measurements

We detect 478 facial landmarks and their coordinates with a facial landmark detection model, a neural network that builds a 3D mesh of the face from photographs. The feature scores are based on the measurements calculated from these coordinates.

Each measurement is compared with a reference range, which can differ by sex. The comparison informs the score of the matching features.

The sections below explain what measurements are taken from the frontal and the lateral photographs:

Frontal photograph

MeasurementReference range
Symmetry95% and above: excellent; 90–95%: good; below 90%: visibly asymmetric
Facial thirdsAbout 0.33 each
Facial fifthsAbout 0.20 each
Canthal tiltWomen: +4° to +8°; men: +2° to +6°
Midface ratioLower values indicate a compact midface
Facial indexAbout 1.30; above 1.45: long face; below 1.25: short or wide face
FWHRWomen: 1.75–1.90; men: 1.90–2.10
Eye spacingAbout 1.0; above 1.15: wide-set; below 0.9: close-set
Jaw-to-cheekbone width0.70–0.82
Upper-to-lower lip ratio0.50–0.66
Nasal indexLower values indicate a narrower nose

Lateral photograph

MeasurementReference range
Nasofrontal angle115°–130°
Nasolabial angleWomen: 95°–110°; men: 90°–100°
Facial convexity165°–175°
Gonial angle120°–130°; above 135°: obtuse jaw angle
Chin projectionPositive values indicate a projected chin, negative a recessed one

Formulas

Measurements are expressed as ratios, so they do not depend on image size. Symmetry is calculated over nine landmark pairs reflected across the facial midline:

mean distance between paired landmarks
cheekbone width
symmetry score; 100 means the paired landmarks match exactly

The facial thirds divide the face height into three spans:

heights of the upper, middle and lower spans
their sum

Width ratios use the cheekbone width as the common denominator, and eye spacing uses the eye width:

jaw width and nose width
face height
distance between the inner eye corners
mean eye width

Canthal tilt is the angle between the inner and outer corners of the eye; a positive value means the outer corner is higher. Full definitions are in How Overchat face analysis works.

Scoring procedure

After the geometry module has mapped the landmark coordinates on both photographs and calculated the measurements, the measurements and photographs are sent to an LLM (large language model), which assigns the overall score and the feature scores in the steps below.

1. Model inputs

The model receives:

  • both photographs
  • the facial measurements
  • sex and age

Sex selects the reference ranges for the measurements and the reference faces for the overall score.

The measurements come in as fixed values, and the model does not estimate proportions or angles from the image on its own. An AI model reading a photograph judges distances and angles only approximately: its estimate can change from one run to the next and can drift toward its overall impression of the face. Measured values avoid this. The same landmarks always give the same canthal tilt or jaw ratio, small differences between faces are not lost, and each feature score can be traced back to the numbers behind it.

2. Feature scores

The model scores 13 features from 0 to 100. Where a measurement exists for a feature, the score starts from it. The rest, such as skin quality and hairline, cannot be measured from landmarks, so the model assesses them from the photographs.

FeatureMain inputs
HarmonyFacial thirds and fifths, facial index, overall balance
MidfaceMidface ratio
EyesEye spacing, eye shape
Canthal tiltCanthal tilt
Skin qualityPhotographs
CheekbonesFWHR, cheekbone prominence
Nose harmonyNasal index, nasofrontal and nasolabial angles
LipsUpper-to-lower lip ratio
SymmetrySymmetry score
JawlineJaw-to-cheekbone ratio, gonial angle
ChinChin projection, facial convexity
Masculinity or femininityPhotographs, FWHR
HairlinePhotographs

Feature scores are read in bands:

ScoreBand
90–100Model-tier
75–89Clearly above average
60–74Above average
45–59Average
30–44Below average
Below 30Marked weakness

3. Overall score

The model then assesses the face as a whole, because overall attractiveness is not the mean of individual features (Rhodes, 2006).

The weight of each feature in this judgment depends on sex, because preferred facial traits differ between female and male faces:

  • Female faces: harmony, then midface, eyes, skin, nose, lips and face shape, then bone structure
  • Male faces: harmony, then bone structure (jaw, cheekbones, chin) and masculinity, then eyes, midface, skin, nose and lips

The face is placed on a scale of reference faces of the same sex, from 3.5 to 9.5. The model names the two closest reference points. The scale runs from 1 to 10 with 5.0 as the average, so half of people score below 5.0 and a score shows distance from a typical face.

Percentile estimation

The model also estimates where the face ranks among people of the same sex. If it leaves the percentile out, the percentile is taken from the score on a fixed normal curve with 5.0 as the average.

overall score, 1–10
cumulative normal distribution
percentile, kept within 1–99
A score of 7.0 is above 91% of faces on the curvePercentile by overall score
1%50%91%99%
12345678910
11%
22%
39%
425%
550%
675%
791%
898%
999%
1099%

Tier classification

The model names the tier. If it leaves the tier out, the tier comes from the score:

ScoreTier
8.5+Exceptional
7.5–8.4High harmony
6.0–7.4Above average
4.5–5.9Solid
Under 4.5Developing
Fewer than 5% of faces reach High harmony or aboveShare of faces in each tier on the same curve
Developing36.9%
Solid37.8%
Above average20.5%
High harmony3.8%
Exceptional1.0%

References

  1. Rhodes, G. (2006). The evolutionary psychology of facial beauty. Annual Review of Psychology, 57, 199–226
  2. Hönekopp, J. (2006). Once more: is beauty in the eye of the beholder? Journal of Experimental Psychology: Human Perception and Performance, 32(2), 199–209
  3. Langlois, J. H. et al. (2000). Maxims or myths of beauty? Psychological Bulletin, 126(3), 390–423

All methodology topics