UPD 29.09: V3 is out.
What's new:
1) It's good at ******** **** (doggystyle, cowgirl) again.
2) The visuals are darker and have more contrast compared to V1 / V2.
3) Also, I've noticed people wanted to create more '*******' looking stuff, so I've added some low quality pics into the dataset. Use "lowres, low quality, candid photo, *******" tags in your positive prompt.