According to forum discussions seen by the IWF, offenders start with a basic source image generating model that is trained on billions and billions of tagged images, enabling them to carry out the basics of image generation. This is then fine-tuned with CSAM images to produce a smaller model using low-rank adaptation, which lowers the amount of compute needed to produce the images.

They’re talking about a Stable Diffusion LoRA trained on actual CSAM. What you described is possible too, it’s not what the article is pointing out though.