Using reference images is a huge step for this sort of thing. The text-only approaches I've seen before were never going to be that good even with "perfect" AI, simply because describing 3D objects in text is not something that anyone is really any good at.