I imagine it would be better at describing an image in a general sense, but probably isn't processing it in a sense where it would actually be counting individual features. I could be wrong about that, but it seems like a combination of traditional CV and an LLM might be what's needed for more precise feature identification.