I just added my training json.

Not the images though. I'm not sure if sharing those will get me in trouble. I can walk you through collecting them though. It didn't take too much time. Maybe an hour or so.

I used imgbrd-grabber to get the images, all from gelbooru and danbooru. I had it save txt files with the tag data, comma separated,
I have seven image sets in my img path. From largest to smallest: cunnilingus, pov_breasts, pov_crotch, futanari_pov, female_masturbation, fingered_pov, fingering_pov
For each of those sets, what I did was, I searched for female_pov (or futanari pov), sorted by score, then grabbed the first 100 or so images that I thought fit the specific concept, like breasts, crotch, etc,
Those 100 images (or less if there weren't enough) go into it's directory with their tag file. Total, between all of the concepts this was aboutt 410 images for this lora.
From there, it's a matter of using something like notepad++ or a python script to search replace in the text files to remove tags you don't want, and to make sure they're formatted correctly. And it's IMPORTANT to add the missing tags for the concept folder. ie, making sure everything in the pov_breasts folder has that as a tag, and so on.

After that, I ran them all through the tagger script, set to 90% certainty, for each of the models it comes with.

After doing that, this was trained at 768,768 using the kohyaa gui for 20 epochs, and load balancing set for about 333 repetitions in each concept. it '3_cunnilingus concept' for that set, because it was a bit over a hundred images.

If you need clarification, feel free to ask.

And if you don't mind sharing your idea; I can see how difficult it would be.

@meterfetish I'm not escaping anything.
Here's an example for 93ece80c791ade358e8f9a5727519e4a from one of my caption files, this is a direct copy paste, no changes, this is the entire file content for of the images caption:
female pov, cunnilingus, fluffydango, makima (chainsaw man), chainsaw man, absurdres, 2girls, artist name, bite markfemale pubic hair, highres, multiple girls, navel, oral, pov, pubic hair, red hair, ringed eyes, yellow eyes, yuri, female pubic hair, looking at viewer

@jangolemestre222 futa only should be pretty easy to get, this one is already trained pretty hard on 'futanari pov' as a tag, so you might want to try that.

In general, you could actually use this at lower strengths and just use the 'pov breasts' tag and you might get what you want.

20 epochs is alot, and I usually set the repetitions to give me a 1k total per concept per epoch; ie, 10 reps for 100 images.

I might be overdoing it, but for concepts (not styles), I get much better results from overtraining and then running the lora at lower strength.

I'm mostly doing that, because I'm including multiple concepts. If I'd trained this one on only cunnilingus, female_pov, then I would have been fine at like five or six epochs.

I didn't have enough data for the female_pov masturbation pose to survive by itself though; in this case, the data for fingering and cunnilingus are propping up the masturbation pose, which doesn't work great at lower epochs.

@ragnaroook that's a good point; full bodied male pov's are kinda rare. I might do that.

I've got work today, but after that, I'll see what I can do.

I'll see what I can do. The issue with that is how small the data set is. Most female_pov images are with futa.