Just a thought... Many of the images being used to train stable diffusion come from getty images.
How do I know?
Prompt: japanese temple garden with blooming sakura and kami
Half of the images generated have something that looks more than vaguely like the grey getty images stamp on it.
Even when you use the negative prompt "text", some form of ghost proof stamp shows up.
Is there a way to train that out of the next version?