This is very much a test model, so don't expect good results outside of a narrow range of input (not saying I've got bad results from that, just saying it hasn't been tested for wide inputs, so I'm expecting it will skew gender female, & that there will be variation between the various "realism" variants of pony in whether it works.
The point was to find a way to consistently generate multiple girls in a mixed orgy scene, without being limited to one specific positioning like the various "multiple girls" lora do (yes, this one is comprised of monster-sex hentai type art, bc that's where I stumbled on one that kinda works).
Pony generates those easily enough, but above 3 girls, it becomes increasingly hard to get Pony v6 to use detailed styles, and at 4 and above, unless the prompt is VERY minimal, faces will invariably be so badly distorted (regardless of negative prompts or embeddings) that they're not salvageable by any amount of refinement (if running on one's own computer, you might in-paint every face, but then you'd be better off inpainting each character separately).
I found a few combos where the faces aren't skewed so badly (and what face distortion there is can be kept within workable "refine in post" limits if you play around with prompt styles and negatives). Initially, I tried the classic "in case + gothic neon + Disney princesses" with all 3 at 0.2-0.6, but "in case" drags the image back to 1-2 subjects, so i dropped that in favour of 0.2 gothic neon and various levels of Disney princesses, and found that worked with the prompt I was using.
I then tried all the variations of pony (except the realism ones - none of those worked, though some do OK with this lora), with various levels of all Disney princesses, and a few minor style combos, and kept what worked (with some pony variants i used other styles instead of ADP).
To train the lora, I used the images that worked the best, and bc the monster hentai genre has tons of fan drawn orgy pics available for public use, I added a bunch of those from a variety of styles, to try and get the lora to recognize that the multi girl, many couples, orgy was the common factor, and to try and minimise the face distortion problem.
It's very much a test - probably closer to alpha than a beta model. If results are promising, I'll increase the training data by a factor of 10, add gay content, and if there's others using this, I'll ask people to help me find large numbers of monster orgy pics of (hopefully) a wide variety of styles. (Oh, and remove the werewolf focus - used that bc it's what Pony variants find easiest to do orgy scenes with, at least if using the on site creator).
If THAT works, the final step will be to try and widen it to more vanilla orgy pics. But I think that's where the hard limits of pony v6 (and all alternatives) will be a problem- it's easier to create monster orgy pics bc you're only specifying the female characters. I suspect there'll be a problem if trying vanilla human orgies, bc that will suddenly double the number of non-clone characters that have to be generated. But if we get to stage 2, and that works, I'll at least make the attempt as stage 3 - it might be that a well- trained stage 2 monster orgy lora provides a workable stepping stone for vanilla human orgy lora.
