How the measurements work

AI Attractiveness Test measures a photo using MediaPipe’s 478 facial landmarks. It does not measure your worth, health or suitability for a cosmetic procedure. There is no scientifically established universal attractiveness scale behind this product.

Quality comes first

We require one visible face covering at least 8% of the image. Estimated yaw must stay within 20°, pitch within 15° and roll within 10°. Mean face-region brightness must be 0.25–0.85, with Laplacian variance at least 60 on the resized image. These thresholds are heuristics and need device testing.

Geometry & reference ranges

Distances are normalized against pupil distance. Symmetry uses 14 corresponding landmark pairs around a pupil-centered face axis. Facial thirds use the top forehead landmark as a proxy, not the actual hairline. Jaw angles are two-dimensional contour estimates. A golden-ratio landing page does not imply a validated golden-ratio attractiveness score.

The combined score

• symmetry: 28% of geometry
• thirds: 18% of geometry
• fifths: 14% of geometry
• fwhr: 16% of geometry
• jawline: 12% of geometry
• eyeSpacing: 7% of geometry
• canthalTilt: 5% of geometry

The combined score is 60% geometry on a ten-point scale and 40% optional AI impression. We clamp it to 2.5–9.5 and round to one decimal. Midface ratio is descriptive only. Client-supplied measurements are validated for shape and range but are not independently remeasured by the server; public statistics are therefore observational and may contain manipulated submissions.

Calibration is not yet validated

The planned target mean of 6.1 and standard deviation of 1.2 are configuration targets, not measured population facts. Distribution mapping is disabled until a representative calibration sample establishes input mean and variance. No percentile or population-ranking claim should be inferred from the band names.

Research & limitations

Published research examines symmetry, averageness, sexual dimorphism and social context. Our numeric weights are product choices, not coefficients estimated from those studies. Image quality, culture, individual preferences and model bias can influence results.

• Little, Jones & DeBruine: Facial attractiveness: evolutionary based research (2011)
• Google MediaPipe Face Landmarker documentation

Processing time

Local model loading depends on your connection and device. The AI request has a 12-second timeout per attempt and one retry. A result in ten seconds is a target, not a guarantee.