When using 'futanari' as a keyword, you often get two futanari's having sex, or some random combination of penis, testicles and vagina on one body.

This lora is an attempt to make that happen less often. Ofc, it's AI it's still going to produce weird noise.

Main keywords are: dickgirl, dickgirl on girl, futa on female.

I notice for strength you need a decently high number to minimize the number of unwanted extra genitalia, something like 0.6-0.8. That will however bias towards lighting and positions in the dataset. Note: Despite a good dataset, albiet lazy labelling, this produces overly exaggerated cum and some other oddities (like comically large penises). I may try training this directly on pony realism at some point, see if that fixes it.

In the meantime you might try putting excessive cum in the negative prompt if you have cum in the picture, and use realcum at some weight as a counterbalance - also don't over prompt penis size unless you want clown sized cocks. This does have some advantages though - easy style/lighting conditions, glistening/sweaty bodies and quite detailed cross sections. If you wanted to use it for just those however, you might get away with 0.5 weight.

Training set was approximate 300 pictures synthetically generated using pony realism, mostly auto-labelled with some additional detail on the internal cross-sections, and using some style loras. Day, night, overcast, dramatic lighting, sunlight, indoors, outdoors, dark, candlelit, street, bedroon, sunset, sunny etc, should trigger lighting conditions well providing mild style effects for the lighting. Aftersex, gaping pussy, in pain, pained orgasm, dazed orgasm should work too, hopefully (although 'in pain' is a little rough)

Will likely be best with pony realism.