>>2771102It helps to understand a little about the training process and quality of data. In short, the AI "learns" how to draw based on images fed in + tagged metadata.
Most of the models used here get their data from the boorus. Visit a random page and notice all the tags associated with the image.
https://danbooru.donmai.us/posts/randomThis is how the AI "knows" how to draw something. Naturally, a lot of them are gonna have watermarks, so in the process of it learning, watermarks are sort of baked in.
If you ask for whatever subject, chances are it learned how to draw with a lot of samples that have watermarks.
However, since watermarks are often tagged, it has some idea of what they are. This is why you can put things like "patreon logo" in the negative prompt and it will not draw it. Of course it's not perfect because tags and learning aren't perfect.
