Support me: https://ko-fi.com/elevatedone
Please post suggestions for future models here: https://civarchive.com/articles/5921/suggestion-list
V1: The model was trained on data mostly composed of squatting poses, so that is the main focus, it can also do lying down
I used r34 for this, if someone has a better dataset I will gladly add it to this model
V2: Still Mostly squatting poses (since those are the most frequent) but can do a lot more. Anal Penetration is still a problem (mainly due to the model not knowing what is anal)
V3 (Final): Thrown out any image board dataset that I had, since those were inconsistent. My dataset was around 500 images generated from PonyXL using AutismMix & T-ponynai3 using this prompt. That's it, and it trained a really nice model, not like the versions prior.
Description
Mostly squatting poses, but can do a bit more.
FAQ
Comments (5)
You can get the best datasets from Gelbooru (much easier using "grabber" when properly configured, it autotags and creates a .txt file based on the tags the image already had on Gelbooru)
yes, I know. And now I see that there are quite a lot of pics there. Dunno what my search terms were before, because it gave me like 20 results.
@ELEVATED You can search "dildo_riding" and get 2k results (probably a bit less with default filters on), or just search "dildo" and manually curate images
@datSato yes, after my comment I did that, but somehow I did not curate it enough. Because I trained with them the model gave me more of the "before dildo riding" or "after" images. So it will take a few days, cause I want a larger dataset and around 15k steps. edit: also I am doing a lot more types and things with LoRA training, since I found that I can run it for 12h/day on GColab, I am doing a lot more.
@datSato update is up