I think you're suggesting that, if it was trained on adult sexual abuse material and images of children, it could composite them in a way that made images depicting CSAM. Doing this would be far different to putting Luke Skywalker on a kangaroo as children an adults look fundamentally different, and without source material it couldn't generate that.