Very cool project ! I fear this might have a pretty high hallucination potential (with current models) the deeper you dig into the base image/context and clicking on potentially unrelated elements in the image. Nevertheless, love the idea.