Who does the baby look like?
Add one photo of each parent and one of the baby. It works for older children too. Front-facing, well-lit photos work best.
Result
Add all three photos to find out
The baby's face slides toward the parent it resembles more.
- Feature by feature
Loading face models (12 MB, first visit only)…
Extra
The whole family
Add grandparents, brothers, sisters, aunts or uncles and see who Baby resembles most out of everyone.
- Up to 8 more people in the same comparison.
- One ranking of everyone, from the closest match down.
- One payment, no subscription. Photos still never leave your device.
Secure payment by Stripe. The unlock is saved in this browser, on this device.
Thank you, payment received. The whole-family comparison is now unlocked on this device.
How the match is worked out
Find the face
A detector scans each photo and boxes every face. If there are several, the largest one is used.
Map 68 points
A second model places 68 points on the face: 17 along the jaw, 10 on the eyebrows, 9 on the nose, 12 around the eyes and 20 around the mouth.
Overall score
A face-recognition model turns each face into a fingerprint of 128 numbers. The baby's fingerprint is compared with each parent's, and the closer one wins. The percentages show how much closer, not how identical: 62% means "nearer to this parent than the other".
Feature by feature
Each face is straightened and rescaled so the eyes are level and the same distance apart. Then the shape traced by the points of each feature on the baby is compared with the same feature on each parent. Eyebrows are measured relative to the eye below, so their height counts too.
What it cannot see
- Colouring. The feature rows are pure geometry, so eye colour, hair and skin tone play no part in them.
- Baby proportions. The recognition model learned mostly from adult faces. Babies have rounder faces and smaller noses and jaws, so they sit fairly far from both parents.
- Pose and expression. A turned head or a big smile changes the shapes, which is why another photo can flip a close result.
- Two measurements. The overall score and the feature rows are worked out separately, so they can disagree.