What Traditional Face Shapes Measure
Traditional face shape categories, oval, round, square, heart, diamond, oblong, describe the two-dimensional silhouette of the face: the outline defined by the hairline, jaw, and widest point. These categories are useful heuristics for hairstyling and eyewear recommendations because they predict the visual balance of shapes applied around the face.
From a face recognition perspective, these categories are quite coarse. Two faces categorised as 'square' might have very different inter-ocular distances, very different nose shapes, and very different midface proportions. The same celebrity embedding could sit in the 'square face' category and match users across a wide range of actual feature configurations within that category.
What AI Measures Instead
A 512-dimensional facial embedding captures far more than the face silhouette. It simultaneously encodes: inter-ocular distance relative to face width, midface proportions (nose length relative to face height), jaw angle and definition, cheekbone prominence, nose bridge width, and many other geometric relationships, each varying continuously rather than falling into categories.
This richer representation explains why people with the same traditional face shape can receive very different celebrity matches. Two 'oval-faced' people whose midface proportions differ, or whose eye spacing differs, will have embeddings in different regions of face space and will match different sets of celebrities.