Why ethnicity-aware scoring is the default, not the exception

If you have ever scanned your face on a generic tool and felt the result was off, there is a structural reason it might have been. Most facial analysis tools, including some that explicitly claim to be "personalised to ethnicity," use Caucasian-population thresholds as a hidden baseline.

The dataset bias is not malice. It is path of least resistance: the largest publicly available morphometry datasets were collected in Europe and North America, in populations that were predominantly Caucasian, and downstream tools quietly inherit those baselines.

Why the same parameter has different baselines across ethnicities

Several parameters have meaningfully different population norms across ethnic groups. A few examples:

• Nasal width ratio. East Asian and African-descent populations have broader average alar widths relative to face width than Caucasian baselines. Using a single threshold scores these populations as having "wide noses" by default, which is not a clinical finding, it is a dataset artefact.
• Skin homogeneity. The Fink chromophore variance scale was originally calibrated on Caucasian skin. Applied uncritically to darker skin tones, it underweights melanin variation and overweights erythema.
• Lip vermilion. Average lip fullness varies systematically by population. A single threshold scores everyone against a Caucasian-leaning median.
• Canthal tilt. Gender-typical ranges differ by ethnicity. The published bands for East Asian populations differ from Caucasian bands by 1 to 3 degrees.

How Facet handles it

Facet keys thresholds to seven ethnicity groups at the math layer: Caucasian, East Asian, African descent, South Asian, Hispanic / Latino, Middle Eastern, and Mixed / Other. The user self-selects during onboarding. The scoring engine then uses the corresponding band for every threshold.

This is not a cosmetic relabelling. The underlying numbers (band edges, gender-typical ranges, percentile bins) are different. A 0.31 nasal width ratio is a 60th-percentile score for one group and a 35th-percentile score for another, and the natural / clinical protocols downstream change accordingly.

What is still imperfect

Honest acknowledgement: the seven-group taxonomy is coarser than reality. Mixed-ethnicity users get an averaged baseline by design, which is better than a Caucasian default but worse than a per-population baseline that does not exist in the literature yet. Several modules (hair density in particular) still lean on Caucasian-dominant references because the alternative datasets are not yet published at quality.

Why this is the default, not an upsell

A common pattern in this space is to add "ethnicity awareness" as a premium feature, or to mention it on the marketing page without operationalising it in the math. Facet's choice is the opposite: the default scan is ethnicity-keyed, and it is impossible to scan without selecting a group. There is no version of the product where you get the Caucasian default by accident.

The reason is simple. A scoring engine that systematically misjudges anyone outside its training population is not a better-built engine waiting for a premium tier. It is a worse engine that should not have shipped.