How old does your face look to FAHR-FaceAge?

A selfie demo of FAHR-FaceAge, the open face foundation model from Haugg, Lee et al., The Lancet Digital Health 2026. Your photo is processed in memory and never stored.

Camera is off.
Press “Start camera”, or upload a photo.

– years (facial age)

Typical error ±5.1 years (MAE on APPA-REAL). Left: the aligned crop the model actually saw.

For fun and curiosity only. This is a research model, not a medical device: it estimates apparent age from a single photo and is sensitive to lighting, pose, make-up and demographics.

What is this?

FAHR-Face is a Vision Transformer trained as a masked autoencoder on more than 40 million face photos (WebFace42M). FAHR-FaceAge is that model fine-tuned on 743,152 age-labelled photos from ten public datasets. In the paper, the gap between face age and chronological age in cancer patients was associated with survival.

This page runs the released FAHR-FaceAge checkpoint on a 4-core CPU server with the same pipeline as the official repo:

RetinaFace detects your face and aligns it by the eyes (exactly one face required).
The crop is resized to 112×112 RGB and ImageNet-normalised.
The ViT-Base encoder plus a linear head regresses age in years.

The survival-risk model (FAHR-FaceSurvival) is not publicly released, so this demo only does age.

Privacy

Your photo goes over HTTPS to this server, is decoded in RAM, run through the model, and discarded. The photo and the estimated age are not saved, logged or used for training. We only keep anonymous counts (page visits, how many photos were taken and whether a face was found), with no IP addresses. The camera image never leaves your browser until you press the shutter.