Identical DNA, Different Faces
Identical twins arise from a single fertilised egg and share nearly all their genetic material. Yet anyone who has spent time around twins can tell them apart once familiar with them, minor differences in mole and freckle placement, subtle asymmetries, slight variations in the relative positions of features accumulate during development in ways not genetically determined. Epigenetic differences and random developmental noise produce measurably different faces from the same blueprint.
From a face recognition perspective, these differences are real and measurable signals. The network does not compare DNA,it compares the actual three-dimensional geometric configuration of the face present in the photograph. If twin A has a fractionally more prominent right cheekbone than twin B, the resulting embedding difference shifts their positions in the 512-dimensional space, potentially placing them in the neighbourhoods of different celebrity embeddings.
How Small Differences Produce Different Matches
The embedding space is high-dimensional and the celebrity distribution is not uniform. A small shift in embedding position, caused by a minor developmental difference, can move a face from the neighbourhood of one celebrity cluster to the neighbourhood of a different one. This does not require a large biological difference; it requires only that the shift crosses an invisible boundary between two celebrity clusters.
Photo conditions also amplify the effect. If twin A's available photos show more profile views and twin B's show mainly frontal shots, their embeddings will reflect both biological differences and the different distribution of photo angles. The most informative comparison uses well-controlled, matched photos from both twins.
What This Reveals About the System
The fact that identical twins can receive different top matches demonstrates that the system is sensitive to genuine micro-differences in facial geometry, not just broad category membership. This sensitivity is what makes biometric face recognition viable, it discriminates at a finer level than human observers typically manage for unfamiliar faces.
Modern deep embedding approaches distinguish twins correctly in most conditions, though their embeddings are closer together than those of unrelated individuals, correctly reflecting the genuine underlying biological similarity. The fact that twins can receive different top results from a celebrity database does not mean the system is confused; it means the system is measuring something real.