Base model occasionally generates censored picture.
This remains true even if you include "censor" in the negative prompts. If you specify "mosaic censorship," there is still a possibility that other unaddressed forms of censorship—such as "bar censorship"—might appear instead.
I suggest that, by default—unless explicitly mentioned in the positive prompts—the generation of any watermarks or censorship should be prevented.
That is because most anime images are censored (if the artist is Japanese, they are enforced by law to censor them). Go to danbooru and search these keywords: "censored" and "uncensored", and you'll see that there are ~600k censored images and only ~180k uncensored. So you would need to post-train the model on an uncensored dataset (which will weaken characters and styles) or use a lora.
This remains true even if you include "censor" in the negative prompts. If you specify "mosaic censorship," there is still a possibility that other unaddressed forms of censorship—such as "bar censorship"—might appear instead.
I suggest that, by default—unless explicitly mentioned in the positive prompts—the generation of any watermarks or censorship should be prevented.
Adding "uncensored" to the prompt usually removes any bar censors/etc, if it doesn't the artist tag(s) used probably have it heavily baked in. In that case, you can weigh up uncensored tag or try something like negpip - https://github.com/Haoming02/sd-forge-negpip which works very well. For watermarks, negpip helps with those as well if negative prompting/etc doesn't.
Somehow the model understand censorship is a part of artist style, so if you use a style you also face the same censorship and signature that the artist use. People on civitai also reported this problem