PCA reveals that faces live on a low-dimensional manifold — eigenfaces are the basis
PCA finds directions of maximum variance in face-image space. The first few eigenvectors (eigenfaces) capture most variation.
Any face ≈ mean face + Σᵢ αᵢ · eigenface_i
With just 6-8 components of a 64-dim space, faces are recognizable. This reveals that human face variation is surprisingly low-dimensional — controlled by a handful of "knobs" (gender, age, expression...).
Synthetic faces avoid privacy issues.