This model replicates the style of the vintage porn studio Rodox. The dataset consists of above 3000 images taken from pornpics.com.
You can try to do hardcore scenes but I believe SD 1.5 is not too good at it, whatever dataset you provide it. I haven't tried ControlNet with it with OpenPose, but it can be something to try out.
The model has however no difficulties with women and men standing next together.
I used a runpod container with a RTX4090 and the EveryDream2 template for this.
Description
Initial version
FAQ
Comments (2)
cool! i tried your VintageKnockers checkpoint but i did not like the extremely large breasts, hopefully this one features those less, and keeps the vintage style / clothes : ) downloading right now!
I know this is unsolicited advice, but as I can appreciate what you are trying to accomplish, I'm hoping I can help you get closer to your goal.
first, the model does not follow prompts well. You should spend more time improving the quality of your captions. With 3000 images, I know this is a daunting task. One idea that has worked for me is to run both blip captions, then wd14 captions, then use a script to combine them. This will give you more varied captions and help reign in the model.
Another idea is that quantity does not equal quality. Be more selective with your images. There are some things SD1.5 does very poorly, so eliminate images that involve things that won't train well and focus on images that can lead to good results. you may end up with half the images you started out with, but as they are of higher quality, at least as far as sd1.5 can train on them, your end result will benefit.


















