Which faces teach a model your age?
Every age estimator is shaped by the photos it learned from. Here are the main public datasets, what their labels mean, and what their licences allow.
Updated 2026-10-18
Why it matters
Two questions decide what a dataset can do. Are its labels the real age of the person or the age people perceive? And does its licence allow a public, commercial service to use it? Most well-known datasets are restricted to non-commercial research, which is why Face-Age uses FairFace.
The datasets
| Dataset | Size | Age labels | Licence, as stated |
|---|---|---|---|
| FairFace | 108,501 images | Age groups judged by annotators (perceived) | CC BY 4.0, commercial use allowed. Used by Face-Age. |
| LAGENDA | 67,159 images, 84,192 people, ages 0 to 95 | Crowdsourced votes (perceived) | “CC 2.0”, exact variant not specified |
| APPA-REAL | 7,591 images | Both real and apparent age | Not stated |
| MegaAge / MegaAge-Asian | 41,941 and 40,000 faces | Age distributions from comparisons (perceived) | Non-commercial research only |
| UTKFace | Over 20,000 images | Exact years, 0 to 116 | Non-commercial research only |
| AgeDB | 16,516 images, 570 people | Exact real age, collected manually | Non-commercial research only, on request |
| IMDB-WIKI | 523,051 images | Real age from birth date and photo date; authors warn labels can be wrong | Academic research only |
| Adience | 26,580 photos, 2,284 people | 8 age groups | Research only, revocable licence |
| CACD | 163,446 images, 2,000 celebrities | Approximate years; authors advise against age estimation use | Academic research only |
Figures and licence wording come from each dataset’s official page, checked on 18 October 2026. In every case the photos themselves keep their own copyright.
What this means for an age test
- A model trained on real ages learns the age on the ID card; one trained on perceived ages learns what people would guess. Face-Age shows a perceived age, because that is what FairFace was labelled with.
- Celebrity photos (IMDB-WIKI, CACD) are often retouched and professionally lit, unlike everyday selfies.
- A dataset can be balanced on paper and still give different accuracy between groups: measure it, as we do on the method page.