it’s essentially iterating on known images

No. It's iterating on the common traits of known images compressed plus lots of randomization.

If you trained a Stable Diffusion model on only pictures of Rwandan people, and asked for an image of “a man sitting on a chair” the man will look vaguely Rwandan.

If you train a model on adult pornography and non-pornography with children and adults alike, it might be capable of generating plausible child pornography.

When you train an AI on CSAM, it produces images that are based on CSAM. Real people were victims in the base material, too. Close e-fuckin’-nough. Real people’s victimization is literally the core of how those images are made.

I've just told you how this is not true.

You seem to have that "all or nothing" mindset in an argument, as if you really didn't like someone, then they should be prosecuted as a rapist, a murderer and an arsonist at the same time. Exaggerating, of course.

Point being that child pornography without real victims is something not contested here and has its own implications. You are trying to argue on something out of reach.