Method

How the estimate works. And where it stops.

Face-Age estimates the age a photo suggests. This page describes what actually runs, what is not validated yet, and what we refuse to compute.

Updated 2026-10-18

What runs on your device

  1. Your browser checks the file: JPEG, PNG or WebP, 15 MB max, 40 megapixels max, 320 px minimum.
  2. The image is redrawn into an in-memory canvas (max 1280 px). Orientation is applied, metadata is not kept.
  3. On your action only, the model files are downloaded from face-age.com (19.7 MB, then cached by the browser).
  4. A face detector counts faces. Zero or several faces stop the process.
  5. The face is aligned on 14 landmarks (eyes, nose, mouth, chin, cheeks, brows) the way the training photos were, then cropped to 224 x 224 px. Four views (two zoom levels, normal and mirrored) are averaged. An age model reads this crop and returns one of 9 age groups: 0-2, 3-9, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+.
  6. Quality heuristics (sharpness, exposure, resolution, face size) are computed on the face region.

Current model: FairFace

Age: the FairFace ResNet-34 model by Kimmo Kärkkäinen and Jungseock Joo (arXiv:1908.04913), weights under CC BY 4.0. It was trained on 108,501 faces from the YFCC-100M Flickr collection, using only photos under Creative Commons licences that allow commercial use. The dataset was built to be balanced across population groups.

Its age labels are perceived ages: three annotators judged the age group of each face and the majority was kept. That is exactly what Face-Age shows: the age a photo suggests, not the age on an ID card.

The original model also predicts gender and ethnic origin. We deleted those outputs from the file itself before publishing it (ONNX export, then our pruning and 8-bit compression). The file served by face-age.com can only return an age group.

Face detection and alignment: Human 3.3.6 (MIT) with the MediaPipe blazeface and facemesh models. Runtime: ONNX Runtime Web (MIT), in your browser.

Accuracy measurements are published below, with their limits.

FileMBSHA-256
fairface/fairface-age-q8.onnx21.3805774a4b847037198612aa6fa7ae055939b32f8063033bc6bd1bffe30082fc15
_astro/ort-wasm-simd-threaded.[hash].wasm (onnxruntime-web 1.30.0)14.243398c10d07d229bd91b364548e130e0e51a8e5704b88c7c083ebbeb78842dee2
human/blazeface.json0.08cd7bbfc078270572beb39f9e5ae67aadbd50b5e67cff37e6d4f6b3ea39312e5f
human/blazeface.bin0.54dc9a97fdc50bc43216554bdd69aa3e7b9361a519ee7bdd996a2f69a98a6f9b72
human/facemesh.json0.10b60ca26f404724f43bd2b1575761d8265180e1f053cc0731caa68462927309e7
human/facemesh.bin1.483826da640b0a3021161605369ee6af293f75d518040355b960ec71a3390c1c0b

Apparent age is not your age

Most age models are trained on photos labelled with the chronological age of the person. They learn a statistical link between image and age, not the judgement a human would make. The number depends on light, angle, expression, make-up, beard, glasses, image compression and on who was in the training data.

The result is not biological age, not a health indicator, not a skin diagnosis and not an attractiveness score. It must not be used for age verification, hiring, insurance, credit or access control.

More in the guides: apparent, chronological and biological age and where the model is less sure.

What we refuse to compute

  • No emotion, ethnic origin, supposed gender or beauty score. The shipped age model cannot produce them: those outputs were removed from the file.
  • No face recognition, no face database, no training on visitor photos.
  • No confidence percentage: we do not publish one until it is calibrated on an independent set.

Validation protocol before launch

  • A licensed evaluation set, independent from training, with no person in two partitions.
  • Chronological and perceived age labels kept separate.
  • Mean and median absolute error, detection failures, stability under light changes and documented gaps between groups.
  • Sample size, composition and uncertainty published. Small subgroups do not support general conclusions.
  • Accuracy per age group (confusion matrix), measured on a set that was not used for training.
  • If accuracy is not credible, we widen the groups or remove the feature.

First measurement, 18 October 2026: not validated

Test set: FairFace validation split (CC BY 4.0, not used for training), perceived-age labels. Criteria were set before measuring.

MeasureTargetResult
Model alone (3,000 faces): right age group within one group-96.5 %
Model alone: exact age group-59.0 %
Largest gap between population groups (adults, within one group)≤ 10 points4.1 points, passed
Full site pipeline (450 faces), ages 20-69, within one group≥ 85 %72.5 %, failed
Full site pipeline, adults, exact age group≥ 50 %30.6 %, failed
Same group when light changes (darker x0.6 and brighter x1.4)≥ 80 %79.3 %, failed

The model itself held up; the way the site cropped the face lost accuracy and made people between 30 and 69 look younger.

Second measurement, 18 October 2026, after fixing the crop: not validated yet

The crop was recalibrated on the FairFace train split, then measured on a fresh sample of 418 validation faces never used before. Thresholds unchanged.

MeasureTargetResult
Ages 20-69, within one group≥ 85 %93.8 %, passed
Adults, exact age group≥ 50 %44.2 %, failed
Same group when light changes≥ 80 %81.4 %, passed
No face found≤ 5 %9.1 %, failed (mostly profile or downward faces)
Largest gap between population groups≤ 10 points4.1 points, passed

Third measurement, 18 October 2026, with face alignment on 14 landmarks: 5 criteria out of 6

The face is now aligned on 14 landmarks (eye corners, nose, mouth, chin, cheeks, brows), the way the training photos were prepared. Fitted on the FairFace train split only, measured on a third fresh sample of 399 validation faces. No face over 70 was left unused in the validation split, so that group is not measured here.

MeasureTargetResult
Ages 20-69, within one group≥ 85 %94.7 %, passed
Adults, exact age group≥ 50 %55.6 %, passed
Same group when light changes≥ 80 %83.3 %, passed
No face found≤ 5 %7.5 %, failed (profile, downward or hidden faces)
Largest gap between population groups≤ 10 points4.1 points, passed

In practice: for an adult, the result is the right age group or the one next to it about 19 times out of 20, and the exact group a bit more than half the time. Full tables, per age group and per population group, and the confusion matrix are in the validation report.

Publisher decision, 18 October 2026

On 18 October 2026, the publisher (NETSOLUTION, François Castellani) decided to remove the PROTOTYPE badge with 5 criteria out of 6 met. Still open: faces not found on 7.5 % of the test images (target 5 %), the 70+ group not measured, and no measurement on real selfies yet. These measurements are in progress and will be published here.

Quality checks are heuristics

Sharpness uses the variance of the Laplacian, exposure uses mean luminance and clipped pixels, framing uses the face height relative to the image. Thresholds are starting points to tune on real test photos. They help you take a better picture; they do not judge your face.