- AI
- Data
- Research
On August 31 the corpus run completed. The corpus now covers 2,681 portraits of notable people from 39 peoples, each read by a vision model in a single call against the site's controlled phenotype vocabularies. It holds 28,756 structured observations of facial and body features, recorded as vocabulary values rather than prose.
The trait pages read from this corpus. The head and face families are dense: eyes were read on 2,423 portraits, lips and mouth on 2,413, nose on 2,410 and skin on 2,681, and every one of the 39 peoples carries enough readings in each head and face family to support a trait page.
The corpus turns a question about a people's appearance into counted observations, family by family. Every number on a trait page carries its sample size beside it.
Under the hood
- One vision-model call per portrait, with a single composed prompt covering every photo-assessable dimension of the 22 vocabularies. The answer returns as structured JSON, one section per anatomical family.
- Each family writes to its own observation table: one row per portrait where that family returned values, linked to the notable-person record and stamped with the vocabulary version and the analysis model.
- Every processed portrait is marked as read, including one where a family returns no values, so each portrait passes through the reader exactly once.
- Selection deduplicates by image, since paired catalog rows for the same people can point at the same photograph.
- Readings by family: skin 2,681 portraits, eyes 2,423, head shape 2,416, jaw and chin 2,413, lips and mouth 2,413, nose 2,410, face proportions 2,410, neck 2,409, head hair 2,397.
- structured observations
- 28,756
- portraits read
- 2,681
- peoples covered
- 39
See it on the site
Related build notes
- Aug 7, 2026The catalog's notable people grow from 13,094 to 23,034A Wikipedia backfill adds 9,792 notable people and lifts covered groups from 292 to 453. Three days later, 15,228 rows carry a photo link.
- May 7, 2026Open pipeline, methodology paper and a DOI: rebuild it yourselfThe code that builds the phenotype dataset is public under Apache 2.0, with a methodology paper and a Zenodo DOI, so anyone can rebuild and cite it.
- Aug 31, 2026Measuring the ruler: 73 of 84 dimensions pass a test-retest studyA test-retest study over 196 portraits scored the head and neck dimensions, wrote the results into the vocabulary files, and made them citable.