Dataset collected from Pixiv.
This lora model depends on the 3.9G NAIFU model.
All pictures showed above were made by AOM2 or anything4.5,but I think it will performed good with other models.
For more precise prompts please download the dataset, if you think your results are not good.
Description
FAQ
Comments (7)
how did you train a sexual position like this?
It's easy. Just collect a bunch of this images and use any LORA training scrips in the similar to dreambooth way. Works fine for any pose not only for anime but also for realistic models.
@desm0nt I suppose you just have to make the image set very very consistent. Adding other positions would mess it up. ty
@slashkey Not really. But in general yes, the dataset should only contain the target pose (but you can vary the angles), ideally not overlapped by anything, not contain angles with ambiguous anatomy (as SD likes to break anatomy, you shouldn't help it), and be relatively even (i.e. so the actress/character doesn't look at the camera in every or almost every picture, otherwise you get 100% head turning to the camera from any angle and a very creepy look).
But in general - collect 50-100 pictures of different characters in a clearly visible target pose and everything will go smoothly.
Got nice results using .4 weight and the following tags: vaginal, pussy, penis, cowgirl_position, girl_on_top, <lora:conceptCowgirl_v10:.4>
Sorry but I cant make it work using any models i have I even used the phrases in the prompt that knxo suggested. It randomly worked like twice out of almost a hundred generations i tried... :/
It is working with 0.4.