Skip to content
Deep Dive

What Is a Similarity Score? How AI Ranks Face Matches by Percentage

A face similarity score shows how close two face embeddings are. Ollie measures the distance between your 512-number embedding and each celebrity's (for unit-length vectors this ranks faces the same way as cosine similarity), then rescales it to a percentage that is easier to read. The percentage is for ranking matches; it is not a probability that you look alike.

Liam BradleySeptember 10, 2026 · 4 min read
Heatmap of pairwise similarity scores
Image: Chamaemelum, Public domain, via Wikimedia Commons

From Distance to Percentage

The raw output of Ollie's similarity search is a distance, the Euclidean distance between two L2-normalised vectors in 512-dimensional embedding space. Two identical embeddings have a distance of exactly 0. Unrelated faces usually sit around 1.4, and the largest possible distance between two unit-length vectors is 2. On its own, this number is hard to read.

Ollie turns the distance into a percentage in two simple steps. First, distances are mapped onto a 0–100 scale, where a distance of 0 is 100% and a distance of 2 is 0%. Then the range where real matches fall is stretched, so the differences between your top matches are easier to see. The ranking never changes; only the display does. The formula was not fitted to human ratings, so the percentage is not calibrated.

What the Percentage Is Not

A similarity score of 87% does not mean there is an 87% probability that you are the same person as the matched celebrity. It means your fingerprint is closer to theirs than to most other celebrities' fingerprints. It is a relative measure for ranking, not a probability.

Because of the stretch, the numbers run high: the closest matches land near the top of the scale, and unrelated faces sit around the middle. Compare your matches with each other rather than reading any single number as a verdict.

Interpreting Your Results

As a rough guide, the higher the score, the stronger the facial similarity, and a clear gap between your #1 and #2 matches means more than the exact numbers. Treat any cut-off, such as "above 75% is strong", as a loose rule of thumb, not a measured threshold.

The ranking is more meaningful than absolute numbers. Your #1 match is the celebrity whose face structure is closest to yours in the full 512-dimensional space. If you upload multiple different photos and consistently receive the same top match, that is a strong signal of genuine geometric similarity. A score that varies significantly across photos reflects variation in the photo conditions, not in your actual resemblance to that celebrity.

Frequently Asked Questions

What does a 90% similarity score mean on Ollie?

It means your fingerprint is very close to that celebrity's compared with the rest of the database. It is a rescaled distance for ranking, not a probability of being that person.

Why does my similarity score change between photos?

Different photos produce slightly different embeddings due to lighting, angle, and image quality variations. The score reflects the embedding distance for that specific photo, not a fixed biological similarity.

Is a higher score always a better match?

Higher scores indicate closer embedding proximity, which generally correlates with more visible similarity. However, the ranking (who is #1 vs #2) is often more informative than the absolute score values.

Try it yourself

Find your celebrity lookalike

Upload a photo and see which celebrities you look most like. Free to try, and your photo is never stored.

Find my celebrity look alike

Related Articles