I prepared 20 images. Finding a good samples which looks similar from angle to depth is one hell of a task) The more the merrier I guess, so if you can prepare more than that - good. I just split gid into frames and choose four that I liked, then put them together in one in square format. That's basically it)

Train Lora like any other. Some say that prompts doesn't matter, but I see the difference, if you sorrowly describe what you see.

Put some non existent tags in prompts, then the algorithm will put everything weird in there.

I suggest finding a high quality captions/gif, cause it affects the style, as you can see at my model with reverse anal cowgirl - https://civitai.com/models/25759/dd-reverse-cowgirl-anal-frames

If you cut all images into 512x512 and then create a canvas in photoshop or crita with 1024x1024 - it's much easier to create a template (cause every frame will be exactly 25% of it).

Hope that helps)

@balmonte47 I meant 20 videos/or gifs, so multiply it by 4 if you counting frames)

Pornhub has gifs from videos, but poor tagging, so you will need to watch lots of videos that won't fit your search request. danbooru has good tagging, but it's anime and gif are rare. Sex dat com has good boards and gifs but poor quality.

The best way is to find good high res videos and save frames from it. But it takes time)

Nonexistent tags is just any combination of letters, that's not an actual word, that can be understood by AI. Like "gdhebwja")

@balmonte47 When you are training Loras you create a dataset with images and text descriptions of them. So adding non-existent prompt it's about that text description.

AI is connecting text (which is basicaly letters combinations) with a bunch of pixels, if explain it in easy way. By adding a non-existent tag in training process you basicaly allow AI to absorb any stuff it doesn't understand in that tag. It prevents further influence on style ot mutant limbs. Not all of them though, but some)

Also that tag can become a trigger word for your Lora. Like a character name, for example.

What resolution are you using? And what model?

@kingofbread couldn't find it with a quick search. I could recomend to use some of the full models, like this or like this. Maybe small loralike model couldn't combine well with other loras, dunno)

@kingofbread Oh, anyhentai should work. could you please try following steps from my tutorial? It's HOW TO chapter in description here, I recently updated it)