Note: For best results, use with SDXL Turbo, or better yet TurboVisionXL.
Use with keywords: upskirt, then optionally 'vagina out' if you want that, then 'rear shot' or 'front shot' depending on whether you want them shot from the front or rear.
Extra keywords like 'lifting up her skirt', 'skirt blowing up in the wind', 'bending over', 'crouching', 'squatting', 'legs up in the air' etc. should also work.
Trained on thousands of carefully selected high definition pantyless upskirt images with full detailed captions I wrote systematically but manually.
It generalizes surprisingly well, using the body of the subject rather than any specific details from the training dataset.
It seems to have learned the actual concept of an upskirt, which is what I was aiming for.
I've been sitting on this one since SDXL came out because there were still a lot of bugs I wanted to fix, but for some reason SDXL Turbo fixes them 95% of the time, probably due to turbo's lack of weird anatomy bugs that misalign nudes by default.
Description
Trained on 1024x1024 images.
I have trained a bucketed one at higher resolution which should be better in theory, but it was significantly more buggy.
FAQ
Comments (4)
Do you have the prompts for the images in the grid?
Was it trained on synthetic data (for example generated by SD1.5) or real pictures?
99% real training + regularization images, but the secret sauce is a dash of a retrain of a synthetic nude generator I made long before text to image existed called NudeGAN.
Only enough for it to understand the concept of what it is to nudify X.
@Tophness can you keep working in this nudegan concept and integrate it inside sdxl txt2img or even img2img ?