This kind of plagiarism detects near-verbatim mashups of small snippets. Has anyone seen similar ideas investigated in image generation? I’m used to seeing pap

This kind of plagiarism detects near-verbatim mashups of small snippets. Has anyone seen similar ideas investigated in image generation? I’m used to seeing papers show pairs as evidence that their models aren’t copying: generated images and the closest images in the training set. But that wouldn’t catch this kind of rephrase-and-mashup pattern’s visual analog. Has anybody looked at that closely?