- London, 102 faces (the set used to build it) r = 0.475 [0.31 ; 0.61]
- Multiracial Americans, 107 faces (never seen) r = 0.409 [0.24 ; 0.56]
What the number r says
r says how closely our ranking follows the ranking of human raters. At 0, no link. At 1, the same ranking. On two sets of faces, we land between 0.4 and 0.5.
The first set: London
102 faces photographed in a studio, rated by 2,513 people (Face Research Lab London Set). Correlation: 0.475, between 0.31 and 0.61. This figure is optimistic: the weights were set while looking at these same faces.
The second set: a blind test
107 multiracial American faces, each rated by 36 to 66 people. No weight and no formula was chosen while looking at them: it is the only test of the scale outside the sample that built it. Correlation: 0.409, between 0.24 and 0.56.
What this does not say
- Geometry explains roughly a quarter of the differences between ratings. The rest comes from skin, hair, expression, lighting, style, and personal taste.
- The agreement holds mostly on women's faces. For the 20 men in the second set, no measurable link was found. Twenty is too few to conclude anything, which is exactly why we publish it with its interval instead of removing it.
- Both sets are studio photos. On real selfies, reliability is still to be measured.
- A group of fewer than 30 faces gets no figure at all: it would mean nothing.
What we refuse to measure
- Ethnic origin: never inferred, never used.
- Emotion: the model can read it, we switched it off.
- Health and weight: no estimate, in either direction.