Filter-resolved bilateral symmetry reveals qualitatively distinct geometric signatures in GAN-generated and diffusion-generated face images. We measure global bilateral symmetry, defined as the pixel-wise mean absolute deviation from mirror correspondence, across 1,000 real faces (FFHQ), 1,000 StyleGAN-generated faces, and 1,000 diffusion-generated faces (SFHQ-T2I: FLUX1 and SDXL), under three independent image-preprocessing filters: raw greyscale, topographic isoline, and edge. StyleGAN faces show modestly higher bilateral symmetry than real faces under all three filters (d = 0.20–0.28, p < 10−4), with results confirmed on a held-out validation sample (d = 0.14–0.40, p < 0.002). Diffusion-generated faces present a markedly different pattern: no significant global luminance symmetry bias, but substantially higher edge-based bilateral symmetry (d = 1.06, p = 2.5 × 10−108). Per-model analysis shows this effect is consistent across FLUX1 (d = 0.984, n = 546) and SDXL (d = 1.035, n = 443) independently, suggesting it is a property of the diffusion process rather than any single model. The edge filter, which suppresses global lighting and pose, is the critical test: its results substantially weaken the principal confounds for whole-image symmetry measurement. Global bilateral symmetry is a simple, parameter-free, training-free feature whose filter-resolved behaviour distinguishes generator classes and may serve as an interpretable component in ensemble detectors for synthetic face images.