>>2771102models are based off on stable diffusion, stable diffusion used LAION dataset which has a ton of fucking images from multiple sources, images that were simply tagged as I dunno "art" by the CLIP thing. It's simply attributing watermarks to artworks as a pattern. These watermarks are often described in the datasets. a shitton of images from patreon are going to have the logo in the same place so it's training is reinforced. Furthermore these models are trained later on specific anime datasets to create mixes and those can be trained more on patreon shit or whatnot.
To make those watermarks appear less you just have to use tags for them in the negative prompt eg "text, signature, watermark, username, artist name"
