You make a very similar argument as @Surdon and my answer is the same (in short, my answer to the other comment is longer):

Yes giving everyone access would be a bad idea. I parallel it to controlled substance access, which reduces black-market drug sales.

You do have some interesting details though:

Training a model on real CSAM is bad, because it adds the likeness of the original victims to the image model. However, you don’t need CSAM in your training set to generate it.

This has been mentioned a few times, mostly with the idea of mixing "normal" children photos with adult porn to generate csam. Is that what you are suggesting too? And do you know if this actually works? I am not familiar with the extent generativ AI is able to combine these sorts of concepts.

As far as I can tell, we have no good research in favour of or against allowing automated CSAM. I expect it’ll come out in a couple of years. I also expect the research will show that the net result is a reduction in harm. I then expect politicians to ignore that conclusion and try to ban it regardless because of moral outrage.

This is more or less my expectation too, but I wouldn't count on the research coming out in a few years. There isn't much incentive to do actual research on the topic afaik. There isn't much to be gained because of the probable reaction of the regulators, and much to lose with such a hot topic.