Stable Diffusion models tend to recreate recognizable facial features from training data rather than generating wholly novel facial structures from scratch