Simulated reference sample

What the age bands mean

Ages route into 18–24 and 25–30 bands. Farkas et al. (2005) studied adults aged 18–30, and both bands use the same young-adult baseline because the paper does not publish a matched split at age 25. Users older than 30 see No published reference for this age range and receive no percentile.

Inputs and method

A seeded generator uses the reported means and variability for young-adult forehead height, nose height, lower-face height, face width, and eye width from Farkas et al. [1]. Each is sampled independently in millimeters and kept above 1 mm.

The generator scores only facial thirds and horizontal fifths. The three vertical measurements are divided by their sum. One fifth of face width is compared with eye width. The runtime uses 2D MediaPipe proxies and estimates trichion because it does not detect a hairline; the source used direct measurements on living people. Those methods are not interchangeable.

Nose width, eye spacing, mouth ratios, canthal tilt, lip ratio, lower-face ratio, jaw angle, and symmetry are descriptive, zero-weight observations. They never enter the simulated sample.

The generator creates 10,000 scores for each selected sex. The pooled option has 5,000 female and 5,000 male draws, an equal split chosen for the product rather than an estimate of population prevalence. The same distribution is used for both supported age bands.

For score s, the percentile is the share of synthetic scores strictly lower than s. Ties do not count as lower. The result is rounded to a whole number and clamped to 1–99.

Limitations

• The cited age support ends at 30, and it does not provide a matched split at age 25.
• Source landmarks do not exactly match the app's image-landmark proxies, so this simulation is not calibrated to the runtime measurement process.
• Its source population and direct caliper methods do not represent everyone. The app uses image-landmark proxies.
• The published table provides separate summary measurements, not a joint distribution. Independent sampling discards correlations.
• The broad score ranges are product-defined. The result is a model output, not a clinical measure or population rank.
• The pooled distribution is equally weighted by sex and is not weighted to population prevalence.
• The sample contains no face images, research-only face datasets, app-user measurements, or biometric templates.

Farkas et al. [2] found that several neoclassical formulas did not represent average proportions. It supports showing ranges rather than one prescription. Langlois and Roggman [3] supplies general averageness background but does not validate a specific thirds or fifths formula.

Internal-data switch

An earlier plan proposed switching to an internal distribution after 500 opted-in analyses per age band. That switch is not implemented: the later privacy decision in SPEC §5.1.1(ix) drops metric_samples and prohibits reference-sample contribution. This code does not collect or accept user metric samples.

Sources

1. Farkas LG, Katic MJ, Forrest CR, et al. J Craniofac Surg. 2005;16(4):615–646. doi: 10.1097/01.scs.0000171847.58031.9e. PubMed.
2. Farkas LG, Hreczko TA, Kolar JC, Munro IR. Plast Reconstr Surg. 1985;75(3):328–337. doi: 10.1097/00006534-198503000-00005.
3. Langlois JH, Roggman LA. Psychological Science. 1990;1(2):115–121. doi: 10.1111/j.1467-9280.1990.tb00079.x.