Where the Number Comes From
Ollie's model turns each face into 512 numbers, scaled so the list has length 1. Think of every face as a point on the surface of a giant sphere. Two points close together are similar faces; two points far apart are different faces.
The raw comparison is the straight-line distance between your point and a celebrity photo's point:
- Identical faces: distance 0.
- Unrelated faces: about 1.4.
- The maximum possible: 2, which never happens with real faces.
That distance is closely tied to cosine similarity, the other common way to compare embeddings. For unit-length vectors, distance² = 2 − 2 × cosine, so ranking by one is exactly the same as ranking by the other. For the maths, see the maths behind your face.
How Ollie Turns Distance Into a Percentage
A raw distance like 0.93 means nothing to most people, so Ollie rescales it in two steps:
- Map distance onto 0 to 100, so distance 0 is 100% and distance 2 is 0%.
- Stretch the range where real matches fall, so the differences between your top five are easier to see.
Neither step changes the order of your matches. They only change how the numbers look. The formula was not fitted to human ratings of resemblance, so it isn't calibrated: an 80% doesn't mean 80% of people would agree you look alike.
Read more: Understanding Your Ollie Results: What the Scores and Rankings Mean
What the Numbers Look Like in Practice
When the Ollie team compared every pair among more than 5,000 celebrities (over 17 million pairs), the scores gave a useful sense of scale:
| Pair | Typical score on Ollie's scale |
|---|---|
| Two celebrities picked at random | About 33% |
| The top 1% of celebrity pairs | 58% and above |
| The closest celebrity lookalikes and close relatives | In the range two photos of the same person usually reach |
Those pair scores average each pair's two closest photo matches; your own results score each celebrity by their single closest photo to yours. You can see the closest pairs in celebrities who look alike, according to an AI.
What the Percentage Is Not
- Not a probability. 87% doesn't mean an 87% chance of anything.
- Not comparable across apps. Every app uses its own model and scale. 85% in one app and 60% in another can describe the same distance.
- Not comparable across photos. A sharp, well-lit photo tends to raise all your scores. Compare matches within one search.
- Not identification. A lookalike score never confirms who someone is.
- Not a beauty score. The model is trained to tell people apart, nothing more.
For more on why scores can mislead, see AI confidence vs accuracy.
Read more: The Math Behind Your Face: How Machines Read Bone Structure
Get new guides by email
How to Read Your Scores
- Look at the order first. Your number one is the celebrity whose photo is closest to yours in all 512 dimensions.
- Then look at the gaps. A clear gap between first and second means more than the exact numbers. Five scores bunched together mean several celebrities are about equally close.
- Then repeat with another photo. A name that stays near the top across two or three good photos is your most reliable lookalike.
For a fuller guide to the results page, see understanding your Ollie results.
Comparing Two Specific Faces
Sometimes you don't want a celebrity; you want to know how alike two people are: you and a sibling, two friends, a parent and child. Ollie's compare faces tool takes two photos, embeds both with the same model and shows how close they are on the same scale. Neither photo is stored.
The same advice applies: use similar photos for both people (same light, same distance, both facing the camera) so you're comparing faces, not cameras.