A note on V2 (Beta): The loss for this training was still not ideal, and I may retrain it at some point. But it does appear to produce roughly in the ballpark of target images. Use a lower cfg like 5.5 to avoid stylization. You can go to 0.9 strength but this lora produces a few anomalies like funny hands or eyes, so you should probably add a face, eye, and hand lora or embedding. Should not accept 'wide shot' and 'closeup' as well as size descriptions of pov cock and nonpov cock and front view (for nearer face), back view (for nearer tail end) and side view, lighting instructions, as well as various positional descriptions. It works, and can produce some nice images, but you will need to regen for complex ideas, and adjust detail, sample and steps to the subject. The example images are somewhat roughly thrown together, so probably not at all the best that can be done.
I may retrain a V2.5 in the future with a faster LR. For now, this will do :)
Main prompt is "POV_Devils3way, mmf, 2boys, 1girl, mmf threesome"
Note on negative prompting:
I added the tag 'semirealistic' to all the synthetic pictures, so if you want to avoid that look, which can be pronounced, put that in the negative tags at whatever weight, and probably if you are getting those you can reduce the cfg or lora weight. Can try something like (cartoon, illustration, semirealistic, sweaty) at some weight. This seems to work!
Double penetration will only work with some picture width, like landscape. But does seem to get an "about to penetrate" double penetration result on portrait, which is new. Faces on dp gens do get a bit messy, so face detailer, or a lora is a good idea.