SD3 was around 10% faster for me then XL and in theory should support up scaling to 3k in some form 2048x3020 or something.For NSFW or in general too? Because I also use AI in general.
It was trained on a far more narrow scope then base XL, and that scope removed or censored in some way NSFW images.
I have tried training just the u-net in diffusers but have been unsuccessful so far. (Not do to memory do to needing a custom script written)
Training U-NET and the XXL TE against Clip may be out of the realm of 24GB card users.
The license basically gives you ownership of the model....unless your using it for commercial means in which case they have it written to sue the hell out of any website....Thus Civatai banned SD3 in less then a week