yeah, images are much harder for copyright I expect

the NYT case is another, but that was an older model more prone to that, and text is easier to (eg.) hash and prevent with a post-gen guardrail