Well their is hope.
F222 model was specifically made for creating nudes, now its been relegated to clothing, so starting with that in your models would be a good first step. Ps the OnlyFakes claims to use the Anything v.3, F222, and "....". I've been unable to dig up what over models they might have used but its likely based off of one of their prompts that they could be entire their own, so don't assume you can make you own personal version of OnlyFakes with ease.
As the the NSFW filter, since the thing is open source people either roll back to a pre-filter build or just code it out how ever they choose.
Last in training, not your training the AI training, theirs several different methods but they all come down to the same thing. You throw it some photos, describe the photos it runs the numbers and spits out what it thinks your were talking about. Your trying to teach the AI to recognize something that wasn't part of the some 2 billion images it was trained on. Be that general things like art styles, or specif wordings, or people that were not part of the training data but you just happen to have a lot of photos of.
At the end of the day though, Stable Diffusion is not snap your fingers easy path to what you want. If you want that then your stuck using someone else premade app and except their limitations and restrictions. And if you want to tune it your self then you need to commit to some hardware. Theirs development to get the systems to be more light weight, and their showing great promises, but in order to actually get your bespoke outputs you are looking at a good $350-$400 graphics card for local running. Their getting it down to the mid 200 ranges as of posting, but its till gonna cost you both money for hardware and time to tune to your desired output.