How a photo becomes an age group.
An age estimator does not “see” a person. It finds a face, lines it up, and compares the pixels with what it learned from thousands of labelled photos.
Updated 2026-10-18
Step 1: find the face
A first, small model scans the image and returns a box around each face. Face-Age stops if it finds none or several: an estimate on the wrong face, or on half a face, would be meaningless. On our test images, 7.5 % of faces were not found, mostly faces in profile, turned down or partly hidden.
Step 2: line it up
A second model places points on the face (eye corners, nose, mouth, chin). The image is rotated and scaled so that these points land where they were in the training photos, then cropped to 224 × 224 pixels. This step matters more than it looks: the model alone gave the right group or the next one for 96.5 % of faces, but only 72.5 % (ages 20 to 69) when our site first cropped faces differently from the training photos.
Step 3: read the crop
The age model is a neural network trained on labelled faces. Ours is FairFace, trained on 108,501 photos whose age group was judged by human annotators. It returns a score for each of 9 age groups (0-2, 3-9, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+). Face-Age reads four slightly different versions of the crop (two zoom levels, normal and mirrored), averages the scores and shows the highest one.
Why a group and not an exact age
The model was trained on groups, so a single number would be false precision. Even groups are not exact: for adults on our test set, the result is the right group 55.6 % of the time, and the right group or the next one 94.7 % of the time. Large independent tests reach the same conclusion: NIST found that accuracy depends strongly on the algorithm, sex, image quality, region of birth and age itself.
Where the calculation happens
On Face-Age, all three steps run in your browser. The model files are downloaded from face-age.com, your photo is never sent. Details and a way to check it yourself: where does your photo go?
Sources
- Kärkkäinen and Joo, FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age, arXiv:1908.04913.
- NIST IR 8525, Face Analysis Technology Evaluation: Age Estimation and Verification, Hanaoka, Grother et al., May 2024.
- Face-Age measurements, 18 October 2026: method and limits.