Are these services using textual-inversion? If so, I have to wonder how well they would work on a stable diffusion model that was trained with the filter in place from the start, so that it couldn't generate anything close to the filter.

As it is right now, stable diffusion can generate adult imagery by itself, however it seems like it's been fine-tuned after the fact to try to 'cover up' that fact as much as they could before releasing the model publicly.

As far as I understand textual inversion != Dreambooth != Actual fine-tuning

I believe the safety filter is trivial to disable since it was added in one of the last commits prior to Stable Diffusion’s public release and not baked into the model, therefore most forks just remove the safety checker code [1]

As far as textual inversion, JoePenna’s Dreambooth [2] implementation uses Textual Inversion.

[1] https://github.com/CompVis/stable-diffusion/commit/a6e2f3b12... [2] https://github.com/JoePenna/Dreambooth-Stable-Diffusion