Methodology: How Nanorater Measures a Face

Nanorater shows its working. This page sets out exactly what the free scan measures, how each number is calculated, what the paid report adds, and where the numbers stop being meaningful.

1. Finding the face

The free scan runs in your browser. Google's MediaPipe Face Landmarker finds 478 landmarks on the face in your photo: the outline of the face, brows, eyes, irises, nose, lips and jaw. Your photo is not uploaded for this step.

Before measuring, the face is levelled: we remove head roll so a slightly tilted photo is measured as if the head were straight. We also check four limits — head roll above 6°, a head turned more than 12% of face width, a face narrower than 18% of the frame, or wider than 80% of it. If any limit is crossed, the result is still shown but flagged as less reliable.

2. The four measured terms

From the levelled landmarks we measure four things:

• Symmetry — how closely the left and right sides mirror each other across the facial midline, using paired landmarks (eye corners, brow points, nostrils, mouth corners, jaw points).
• Facial thirds — how the face divides vertically into upper, middle and lower sections.
• Facial fifths — how the face divides horizontally into five widths across the eyes.
• Proportion — the face's length-to-width ratio.

Each term is a distance from the typical value of a reference set: 1,088 level, frontal photos from a public face dataset, measured exactly the same way. For thirds, fifths and proportion the typical value is the reference set's median; for symmetry it is a perfect mirror.

3. From distance to a 0–100 score

Each distance becomes a closeness score:

score = 100 × exp(−distance ÷ scale)

The scale is chosen per term so that the reference set's own median portrait scores 80. A score above 80 means your photo sits closer to typical than the median reference photo did; below 80, further away.

The overall face score is a weighted average: symmetry 30%, thirds 25%, fifths 25%, proportion 20%. These weights are Nanorater's own calibration, not a scientific constant.

Next to every term we show a percentile — the share of reference photos your photo measured closer than — and a typical range, the middle half (25th to 75th percentile) of the reference set.

4. Guide scores: golden ratio, eyes and jaw

The harmony profile adds three dimensions that have no reference-set calibration, so we label them guide scores:

• Golden — eye spacing against the 0.46 ratio research associates with attractive faces, and eye-to-mouth distance against the reference median.
• Eyes — canthal tilt (the angle from inner to outer eye corner) against a 5° ± 3° guide.
• Jaw — lower-jaw width relative to cheekbone width against a 0.80 ± 0.05 guide. The frontal jaw angle is shown in the report but not scored.

Guide scores use the same closeness curve, scaled so one spread away from the target scores 80 (for the golden ratios, the spread is half the reference set's interquartile range).

5. Face shape

Face shape comes from a small statistical model (multinomial logistic regression) trained on 2,497 labelled photos from a public face-shape dataset, using the widths of the forehead, cheekbones, jaw and chin relative to face length. Tested on photos it never saw, its first choice matched the label 50% of the time and its top two 72% of the time; guessing would score 20% and 40%. That is why we always show a ranked list, and say "between two shapes" when the model is unsure.

6. PSL-style estimate

The PSL-style number is a direct mapping of the overall score onto a 1–8 style scale: a score of 80 (the median reference portrait) maps to 4, 100 maps to 6, and 0 maps to 1, rounded to the nearest half point. It is a translation of the same measurement, not a second opinion.

7. What the Full Report adds

The paid Full Report starts from the same measurements. Your photo and the measured numbers are then read by a vision language model, which writes region-by-region readings (eyes, brows, nose, lips, jaw, skin, hair), a ranked improvement plan, a four-week schedule and styling guidance. It is instructed to stay with hair, grooming, skin care, styling and photo technique — never surgery, procedures or diagnoses — and to say when a front photo cannot support a claim.

Four AI previews are then generated on your own photo: a hairstyle, grooming, better photo light, and a styling look. They are visualisations of styling changes, not a forecast of how your face will change.

8. What the numbers cannot tell you

• One photo is not a face. Angle, lens distance, light and expression all move the measurements. Use a level, front-facing photo at arm's length in even light for the fairest read.
• Typical is not ideal. The reference set is a disclosed yardstick of celebrity-style photos, not a picture of the world or a standard of beauty.
• Attractiveness is more than geometry. Skin, hair, expression, style, voice and personality are not in these numbers.
• It is not medical. Nothing here is a diagnosis or a reason for a procedure.

Read more on our Limitations, Bias & Fair Use page.