We should start looking into training a Flux lora. We'd need to build a new dataset with captions. Then we'd need someone with a 24gb vram card to train it, or we'd rent compute somewhere. Renting is just a few bucks.
We might need to wait for someone to create a good nsfw model first thought. Otherwise genitals will be bad.
The advantage of training on flux would be generations with fewer errors and more complex scenes, including readable text.
The buzz cost for flux generations on civitai is much higher however, but you can run flux locally even with as low as 4gb vram. However, with such low vram, you need to lower the resolution.
That being said, I don't really feel like building a SS dataset lately. This is something the community should do collectively, though.