> But generating a model-image that suits a given demographic, that expresses "what's becoming hip right now" and so-forth would be extremely hard

Just generate a spread of different looks, test them on different audiences, and measure engagement. Feed that back into the parameterization for generating the next set of images.