This isnt as interesting as it appears, its just feeding the signals we could already get into the ai then hallucinating a completely new but similar image.

Now if you were to instead train the AI on the signals and the GT images instead of using someone elses model we might get somewhere interesting