Skip to content
Back to home03 / How scoring works

How Our Face Rater Calculates a Score

Follow the measurements, six dimension weights and worked examples. See what a score can tell you and what it cannot.

AuraFace finds facial landmarks in one front-facing photo, measures ratios, then applies published rules to calculate a reference score. The AI model locates points. The rules produce the score. The system is not trained to predict the rating other people would give your face.

478 points do not mean 478 scoring rules

MediaPipe Face Landmarker estimates positions around the eyes, nose, lips and facial outline. AuraFace uses those positions for 17 geometric rules, grouped into six dimensions.

DimensionWhat is measuredWeight
Face proportionsFace height relative to cheek width, and the upper face segment relative to face height.20%
Visible symmetryDifferences in eye width, eye and mouth level, cheek distance from the centre line, and nose-tip position.20%
EyesInner-eye distance relative to eye width, eye width relative to face width, and eyelid opening.15%
NoseNose width relative to inner-eye distance, and nose length relative to the central face segment.12%
MouthMouth width relative to face width, and lower-to-upper lip thickness in the photo.14%
Lower faceJaw and temple widths relative to cheek width, and the mouth-to-chin segment relative to face height.19%

Temple width and face height use fixed model landmarks. They are measurements in the photo, not head circumference, a precisely detected hairline or clinical measurements. Skin texture, expression in motion and nose-bridge height in profile are not scored by these six dimensions.

How does a measurement become a score?

Take rule C1, face height divided by face width. These numbers are a calculation example, not measurements from a user.

  1. Suppose face height is 220 pixels and cheek width is 160 pixels. The ratio is 1.375. Doubling both measurements gives 440 ÷ 320, still 1.375.
  2. The current C1 curve has its centre at 1.30. Its internal value falls smoothly as a ratio moves away from that centre. This is a product rule setting, not a universal ideal face shape.
  3. Using the current curve, 1.375 produces an internal C1 value of about 91.7/100. That is one rule, not the composite score or an appearance percentile.
Show the formula for this example

This example is above the curve centre. Its width parameter is (1.75 − 1.30) ÷ 2.5 = 0.18. The internal value is calculated as follows.

100 × exp(−0.5 × ((1.375 − 1.30) / 0.18)²) ≈ 91.7

Other ratios have their own curve parameters. Symmetry rules start from zero visible difference. The rule source below lists the parameters.

How do the dimensions combine?

Valid rules are combined within each dimension, then the weights above produce an internal total. That total is mapped to the report’s 10-point scale. For a hypothetical example, use these internal values:

Face proportions
94
Visible symmetry
92
Eyes
91
Nose
93
Mouth
90
Lower face
92

The weighted internal total is 92.09/100. With the current display mapping, that becomes 6.2/10. This illustrates the calculation; it is not a human rating study.

Each displayed dimension has its own scale mapping and rounding. Averaging the six displayed scores will therefore not necessarily reproduce the composite. If measurements are missing, the remaining rule coverage is checked before a total can be shown.

6 out of 10 does not mean above 60% of people

The display scale is calibrated using 20,000 synthetic geometry samples. These are generated parameter combinations, not attractiveness votes for 20,000 real people. The distribution chart explains this product’s scale, not a population ranking.

Photo checks decide whether measurement is usable

Passing the photo checks means the image is suitable for measurement. It does not add points. An unsuitable photo can lead to a retake prompt or a report without a composite score.

  • People and framing: Use one complete face. Keep the chin and sides inside the image. A face that is too small loses useful detail.
  • Pose: Turning changes visible left-to-right distances. Tilting can also affect ratios. Current pose estimates mainly use 2D landmarks and cannot precisely recover all 3D head angles. A passing check does not remove every perspective effect.
  • Sharpness and exposure: Motion blur, backlighting and blown highlights make landmarks harder to locate. Start with a clear original rather than adding sharpening or beauty filters.
  • Expression: A broad smile, pursed lips or an open mouth changes eye opening and mouth measurements. Expression checks currently warn rather than reject on their own. A relaxed expression makes separate results easier to compare.

Why can the same person receive different results?

The inputs are projected proportions in a photo. Distance, lens, expression and landmark error can change those inputs. To check a result, use the same camera, a similar distance and even lighting. The current system analyses one photo at a time; it does not merge several photos into a stability-checked result.

Read the photo feedback before the number

  1. Check photo quality first. Retake a blurred, turned or cropped photo before interpreting facial measurements.
  2. Use the dimensions to locate the main differences. An eye dimension covers a few geometric ratios, so a low value is not a complete judgement of how your eyes look.
  3. Open the detailed analysis to read the measurement, rule and validity. A ratio away from a product rule centre does not automatically mean it needs changing.
  4. Then consider styling suggestions. Pick a goal you care about, such as showing more forehead or keeping width at the sides. These are options to try with hair or frames, without a promised score increase.

An explainable calculation still has limits

We do not have an independent human-rated benchmark establishing how closely these rules match public preferences, so we do not publish an attractiveness accuracy percentage. Landmark detection, geometric measurement and aesthetic preference are different tasks. Published parameters make a calculation inspectable; they do not make its preferences universal.

The system does not assess skin condition, age, personality, health or ethnicity. A single frontal photo cannot establish a complete skeletal structure or profile. Faces with similar scores can look different because the rules cover only part of what people see.

Where are the photo and report kept?

Your photo stays in your browser and is never sent to AuraFace servers. Measurements run locally. Your report stays in this browser session, and you can download a copy to your device.

The browser still downloads the model and website resources. Local photo processing does not mean there are no network requests, or that the report is saved permanently. Use the download option on the result page if you want to keep it.

Implementation and references

This explanation follows the current implementation. Reports from different rule versions may not be directly comparable. Current scoring version: local-rules-3.2.0.

Try a photoPhoto and result questions