Why Some Celebrities Match More People
In any face database, some embeddings sit in densely populated regions of the embedding space while others sit in sparse regions. A celebrity whose embedding falls near the centroid of the distribution, mathematically average across many facial dimensions, will be close to more people's embeddings than a celebrity with unusual proportions. This is a purely geometric consequence: a point near the centre of a cluster is, by definition, close to more other points than a peripheral one.
The most-matched celebrities are therefore not necessarily the most famous ones, or the most photographed ones, or the most attractive ones. They are the ones whose facial geometry sits closest to the average of the human faces their embedding space was built to serve.
The Averageness Effect
Psychologists have long known that mathematically averaged (composite) faces are rated as more attractive than individual faces. The same principle applies in face matching: faces near the mathematical average of the embedding space also attract more matches. A celebrity with face proportions close to the global mean across multiple dimensions will consistently appear in more top-5 results than a celebrity with distinctive outlier proportions.
This creates an interesting pattern in Ollie results. Users who receive a match with a very commonly matched celebrity tend to have facial proportions in the most populated region of the embedding space, their face is broadly representative, not unusually distinctive. Users who receive matches with rarely-matched celebrities tend to have more distinctive proportions that only overlap with a small set of similarly configured faces.
What This Reveals About Your Results
If the same celebrity consistently appears at the top of your results across multiple photos, they are almost certainly the nearest neighbour to your true facial embedding in the high-dimensional space. This is a reliable signal regardless of whether that celebrity is commonly or rarely matched.
If your top match shifts between photos, you are likely in a region of the embedding space where several celebrity embeddings are at similar distances. Small changes in photo conditions (lighting, angle) shift your query embedding slightly, moving the closest celebrity neighbour. Looking at the top 3–5 results across several photos gives a more complete picture of which celebrities genuinely cluster near your face.