TheFaceRating

How the face rating works

The complete method behind every free tool: which landmark model we use, how a face is levelled and calibrated, the 11 proportion formulas and their weights, and what a single photo can't tell you.

1. The landmark model

Every tool starts with Google's MediaPipe Face Landmarker, a neural network that predicts 478 three-dimensional points on a face from a single image: 468 for the face surface (Kartynnik et al., 2019) plus 10 for the irises (the attention-mesh refinement, Grishchenko et al., 2020). It runs inside your browser through WebAssembly, on the GPU where available and on the CPU otherwise. The model file is downloaded from Google's servers; your photo is not sent anywhere.

MediaPipe returns coordinates normalized to the image's width and height. We first convert them to true pixel space so that a portrait or landscape photo can't stretch any ratio.

2. Pose checks and head-roll normalization

Before measuring, the scanner rejects photos that would give misleading numbers. A head turned to one side is detected from the nose-tip distance to each cheekbone: if |left − right| ÷ (left + right) exceeds 0.14, you're asked to face the camera. Very dark or bright photos, a face that fills too little of the frame, an open mouth or a smile are flagged too, because each one moves specific landmarks.

Head tilt (roll) is corrected rather than rejected. We measure the angle of the line through the two iris centers and rotate every point around the brow center until that line is horizontal:

θ = atan2(y473 − y468, x473 − x468)
p′ = rotate(p, around L168, by −θ)

All ratios, angles and the symmetry midline are measured on these levelled points.

3. Calibration to the reference scale

Classical norms (Farkas, 1994) measure from true anatomical points: the hairline, the brow line, the jaw corner. MediaPipe's nearest landmarks sit slightly inside those points, so raw mesh ratios read every face as shorter, broader and squarer than it is. Five ratios are therefore mapped linearly onto the reference scale, reference = refMean + (raw − meshMean) × spread, using these constants:

RatioMesh meanReference meanSpread
Length ÷ width1.21.4261
Forehead ÷ cheekbones0.730.831
Jaw ÷ cheekbones0.8850.82
fWHR2.151.881
Jaw angle (°)1371251

The mesh means come from MediaPipe's canonical face model and a small set of early scans; the reference means from published norms and our celebrity reference set. Both will be re-fitted as more data becomes available. Raw values are stored with every scan so results can be recomputed.

4. The 11 proportions and the harmony score

Each proportion is scored 0–100 by how close it is to its reference value, using a Gaussian curve: score = 100 · e^(−½ · ((value − target) ÷ tolerance)²). A measurement exactly on target scores 100; one tolerance away scores about 61; two tolerances away about 14. Symmetry and thirds use their own exponential curves, shown below.

ProportionFormula (MediaPipe indices)Target ± toleranceWeight
Bilateral symmetry100 · e^(−3.2·D), D = mean mirrored discrepancy of 12 landmark pairs ÷ ½ cheekbone width88–100 band20%
Facial thirds balanceΣ |third − ⅓| over 10→168, 168→2, 2→152 (each ÷ 10→152)100 · e^(−dev ÷ 0.09)14%
Canthal tiltatan2(y_inner − y_outer, |x_outer − x_inner|), mean of both eyes+5.5° ± 4°10%
fWHR|234 − 454| ÷ |168 − 0| (calibrated)1.90 ± 0.188%
Length-to-width|10 − 152| ÷ |234 − 454| (calibrated)1.45 ± 0.1410%
Midface ratio|468 − 473| ÷ vertical pupil-line → 01.00 ± 0.0810%
Eye spacing|468 − 473| ÷ |234 − 454|0.46 ± 0.038%
Jaw-to-cheek|58 − 288| ÷ |234 − 454| (calibrated)0.80 ± 0.097%
Nose width|102 − 331| ÷ |133 − 362|1.00 ± 0.165%
Mouth-to-nose|61 − 291| ÷ |102 − 331|1.55 ± 0.224%
Jaw angleangle 234–58–152, mean of both sides (calibrated)125° ± 14°4%

The weighted average W of the 11 scores becomes the harmony score, clamped to 35–98 so a single photo can never claim a perfect or a minimum result:

harmony = clamp(21 + 0.68 · W, 35, 98)

Benchmark percentiles on tool pages are modelled, not observed: a score is mapped through a normal distribution with mean 60 and standard deviation 18. We don't build a database of user faces, because selfies are deleted within hours and free results after 48 hours.

5. Face shape

Face shape uses three calibrated ratios: length ÷ cheekbone width, forehead ÷ cheekbone width and jaw ÷ cheekbone width. We compute a weighted distance to six prototypes (oval, round, square, heart, diamond, oblong) taken from the average ratios of our celebrity reference set, add small priors for unambiguous cases (for example, length ÷ width of 1.55 or more favors oblong), and turn the distances into a probability for each shape. That's why results show a main shape and how close you are to the others.

6. What runs where

7. Limitations

Every tool page shows the specific formula, landmarks and reference bands it uses, plus how to take the same measurement by hand. Start with the free tools or read our editorial policy.

Methodology questions

Is my photo sent to a server for the free face rating?

No. The MediaPipe model and all the formulas on this page run in your browser. Only the numbers are saved when you open your results page; the photo stays on your device unless you separately consent to the AI color analysis.

Why does my score change between photos?

Every ratio is measured on a 2D projection of a 3D face. Camera distance, a slight head turn or tilt, expression and lighting move landmarks by a few pixels, which moves the ratios and therefore the score by a few points. Use several front-facing photos taken from about a metre away and look at the range.

Are the percentiles based on other users' faces?

No. We don't keep a database of user faces. Percentiles are modelled: each 0–100 score is mapped through a normal curve (mean 60, standard deviation 18). They show where a score sits on the scorer's own scale, not a ranking against real people.

Why do you calibrate the MediaPipe measurements?

MediaPipe's landmarks don't sit exactly on the classical anthropometric points: landmark 10 is below the hairline and 58/288 sit above the true jaw corners. Measured raw, faces read shorter, broader and squarer than they are, so five ratios are mapped linearly onto the reference scale before scoring. The raw values are kept with every scan.

Is the harmony score a measure of attractiveness?

It measures how close one photo's proportions are to classical reference values. Attractiveness research shows that proportions, symmetry and averageness play a part, but so do expression, skin, grooming, context and the person looking. Treat the score as a description of geometry, not a verdict.

References

  1. Kartynnik Y, Ablavatski A, Grishchenko I, Grundmann M (2019). Real-time Facial Surface Geometry from Monocular Video on Mobile GPUs. CVPR Workshop on Computer Vision for AR/VR.
  2. Grishchenko I, Ablavatski A, Kartynnik Y, Raveendran K, Grundmann M (2020). Attention Mesh: High-fidelity Face Mesh Prediction in Real-time. CVPR Workshop on Computer Vision for AR/VR.
  3. Farkas LG (1994). Anthropometry of the Head and Face (2nd ed.). Raven Press, New York.
  4. Rhodes G (2006). The evolutionary psychology of facial beauty. Annual Review of Psychology 57:199–226.
  5. Thornhill R, Gangestad SW (1999). Facial attractiveness. Trends in Cognitive Sciences 3(12):452–460.
  6. Little AC, Jones BC, DeBruine LM (2011). Facial attractiveness: evolutionary based research. Philosophical Transactions of the Royal Society B 366:1638–1659.
  7. Todorov A, Olivola CY, Dotsch R, Mende-Siedlecki P (2015). Social attributions from faces: determinants, consequences, accuracy, and functional significance. Annual Review of Psychology 66:519–545.
  8. Willis J, Todorov A (2006). First impressions: making up your mind after a 100-ms exposure to a face. Psychological Science 17(7):592–598.