Geometric Position in Embedding Space
The primary factor determining how often a celebrity gets matched is the position of their embedding in the high-dimensional face space. Celebrities whose embeddings fall in dense, central regions of the space are geometrically close to many user face embeddings and will appear frequently in results. Celebrities in sparse, peripheral regions will appear rarely, but when they do, the match is usually visually strong.
This geometric explanation is separate from popularity. A highly famous celebrity with distinctive, unusual proportions may appear less frequently in results than a less famous celebrity with average proportions, because the unusual proportions place their embedding in a sparse, peripheral region with few nearby faces.
Dataset Representation
Training data quantity and diversity also influence embedding quality. Celebrities with many varied training images, photographed across many years, conditions, and contexts, have well-positioned, stable embeddings. Celebrities with fewer training images have embeddings derived from a narrower distribution of conditions, which may be less representative of their true face geometry.
Ollie's network was trained on MS1MV2, which covers tens of thousands of people photographed in many different conditions. But dataset asymmetry is unavoidable at scale: more prominent celebrities simply have more available photographs across more varied conditions.
Distinctiveness as a Two-Edged Sword
Distinctiveness affects matching frequency in two ways. Highly distinctive celebrities, those with unusual or extreme proportions, appear in fewer results because fewer users' faces land near them in the embedding space. But when they do appear, the similarity is often visually obvious: distinctive proportions match only highly similar faces.
Average-featured celebrities are the inverse: they appear frequently, but the matches are sometimes visually subtle. A match based on being near the centroid of many dimensions simultaneously may not produce an obvious visual resemblance, even though the embedding distance is short.