If you run stable diffusion on your local system, the sky's the limit. The official Stable Diffusion models 1.5 and older are completely unfiltered, having been trained on
everything from the LAION 5B data set. The official Stable Diffusion 2.0 model is much better for general use, but the authors went to great lengths to remove a lot of material from its training set. Models using or based on 2.0 have far fewer famous people/celebrities and have been stripped of many "styles" based on specific artist names. I wouldn't be surprised if they also removed some NSFW content or keywords. It's no secret that they're trying to "legitimize" their models and are trying to position it in a way that appears to sidesteps some potential copyright concerns.
But the official SD models aren't the
only models.
People are already training new models from scratch or have
retrained the original SD models using additional content, almost all of which fall into that
very grey copyright legality. I've seen models trained on specific niche content like Disney screenshots, Fortnite characters, and Pokemon pixel art to whole catalogs of images from ArtStation, DeviantArt, and the
very NSFW Danbooru.
If you're running your own software, you can download and use and mix any of these models together for interesting results. There are over a thousand "diffuser" models currently hosted on Hugging Face (
https://huggingface.co/models?library=diffusers), but a model is just a big file that can be hosted and shared anywhere else more privately. So, some forums and Discords privately distribute models, and some require paid memberships through services like Patreon.
It's safe to assume that Midjourney is also running its own custom model. It's hard to know if they based in on an older version of SD or trained it from scratch, but they clearly aren't interested in sharing their secret sauce recipe.