Ahem, yes, the hash is the same as the other model, I'm not supposed to mention it, tho.

This model is NOT perfect and will need more training, but has quite good nudity representations, works best with the tags of anime boards.

I noticed that the dynamic thresholding works somewhat good with sexual acts but ruins the nudity.

It is not trained on many sexual acts and they come up pretty bad because Colab disconnected before the end of training and I was training with constant scheduler.

Seems to work nicely with euler A at high CFG and normal CFG of SDE M.