I think they’re saying that it’s not generating slop from nothing. They take the artist’s “structure data” as a “ground truth” and the generation is “guided” to generate slop that won’t deviate too far from the original?