Methodology — Face Rater

Face Rater's free analysis is a deterministic geometry calculation, not an AI opinion. The scoring formulas are deterministic for the same detected landmarks. Landmark detection can vary with browser or model versions. This page documents exactly what is measured, how scores are derived, and where the method's limits are.

The free tools below run locally. A paid report is a separate consented flow: the selected photo is uploaded after checkout for safety review and seven styling concept images. Those images are a visual deliverable, not part of the free score or its measurement formulas.

The pipeline

1. Detection. Your photo is analyzed locally in your browser by MediaPipe Face Landmarker, which returns 478 facial landmarks as normalized coordinates. The original photo never leaves your device in the free tier.
2. Quality gates. We refuse to score a photo when no face is found, when more than one face is found, when the head is turned too far from the camera (estimated from the left/right cheek-to-nose distance), or when the image is too blurry. A slight turn is still scored, with a warning that it can lower the result. No face, no forced score.
3. Preprocessing. Head roll (tilt) is estimated from the line between the outer eye corners and corrected before scoring. Normalized coordinates are converted to a consistent pixel scale — raw normalized x/y values are never compared directly, because x scales with image width and y with image height.
4. Scoring. Three sub-scores are computed and combined into a 0–100 reference score.

What we measure

Symmetry (40% of the score)

• Mid axis: estimated from a combination of the midpoint between the eye centers, the nose bridge, and the philtrum — never from the nose-tip x-coordinate alone, which is easily skewed by pose.
• Compared pairs (at minimum): outer eye corners, inner eye corners, cheeks, brows, and mouth corners. Each pair's deviation from a perfect mirror across the mid axis is measured as a percentage of visible face width.
• Score mapping: 100 × clamp(1 − mean deviation percentage / 8, 0, 1). A deviation of 0.63% produces approximately 92; 8% maps to zero. This is our chosen scale, not a population percentile.

Visible face proportions (35% of the score)

• Width-to-height: visible face width ÷ visible face height, where height runs from the top-of-face landmark (MediaPipe point 10) to the chin (point 152). Point 10 is not the hairline — we do not claim to measure classical "three-section" facial thirds, only the face visible in the photo.
• Segment ratios: brow-to-nose-base and nose-base-to-chin distances relative to the middle segment, using available landmarks.

Eye-area harmony (25% of the score)

• Inner-canthal distance ÷ visible face width — how far apart the inner eye corners sit, relative to the face.
• Eye-width symmetry — the percentage difference between left and right eye widths.

The dedicated face symmetry test exposes the paired deviations and the fitted midline directly. The golden ratio face calculator draws three supported 2-D ratios and compares them with 1.618 as a reference only. No ratio is a universal facial target, and deviation is not a defect.

From measurements to the score

Symmetry maps deviation onto 0–100. Proportions combine the width/height band score (60%) and upper/middle and lower/middle band scores (20% each). Eye harmony combines spacing (60%) and eye-width asymmetry (40%). Band scores decay linearly outside the reference interval. The overall score is a weighted sum — 40% symmetry, 35% proportions, 25% eye-area harmony — mapped to 0–100, then assigned a tier:

• 85–100 Harmonious
• 70–84 Balanced
• 55–69 Developing
• below 55 Starting Point

These tiers describe photo geometry bands, not people. "Developing" and "Starting Point" describe how far the photo's measurements sit from our reference bands — often a lighting or angle artifact — and are not judgments of anyone's face.

Reference ranges: initial values, pending calibration

The reference bands are initial, deliberately wide values. They have not yet been calibrated against a licensed, diverse sample set, and they are not scientific facts or population norms. Until calibration is complete, every score display carries this caveat. If calibration shifts the bands, we will update this page and say so.

Stability rule

The upload quality gates reject unsupported face counts, excessive pose and blur before scoring. These checks do not guarantee that another photo gives the same score. Our controlled photo experiment is documented in the photo conditions guide; it is not a population calibration or a guarantee of stability across devices.

Known limits

• One front-facing photo. Profile features (nasal projection, chin projection, forehead slope) are invisible from the front and are reported as "not scored — needs a side photo," never guessed.
• Photo conditions move the numbers. Lighting direction, head pose, lens focal length, camera distance and device can shift detected measurements. We do not promise a fixed number of points.
• 2D projection of a 3D face. A photograph flattens structure; geometry on a photo is not geometry on a face.
• Not medical. Nothing here is medical, dermatological, or cosmetic-surgery advice.
• No trait inference. We never infer intelligence, trustworthiness, competence, personality, age, or health from a photo.

Ten free readouts and the 35-reading report appendix

The free result shows ten headline measured values alongside three scored components. The expanded report explains 35 photo-space readings from the same landmark data: six left/right pairs each contribute mirror, vertical and lateral differences (18 readings), plus 17 ratios. These sets overlap. The paid report does not add 35 independent beauty scores or a new diagnostic scan. The expanded readings are computed locally for supported scans; the paid value is their interpretation, styling comparisons and report delivery. Older reports may lack expanded readings; unavailable measurements remain unavailable.