My first attempt at training my mooning / upskirt concepts for SD3.

I've only been able to get this model to work in the latest version of Automatic1111.

ComfyUI produces strange blobs, Forge fails to load it at all. This should be fixed soon when they update it to support the newer model architecture.

I found best results using Euler at 50 steps / 3 CFG. It's very hit and miss because it's SD3, the trainer I'm using isn't great, and it's still training.

I also have an idea to completely fix the NSFW censorship and anatomy problems in SD3 coming soon. It's looking like it may take weeks to finish this training run, and in the meantime I have to do some final touches on 56,000 images which will take a long time, but so far the experiment seems to be working.

Before diffusion models existed, I created a model called NudeGAN which would take in any clothed photo and convert it to nude, and luckily I still have the dataset which contained thousands of perfectly consistent 4k images of thousands of women from every angle with the same camera settings, poses and backdrops with and without clothes on.

These are all perfectly labelled in such a way that it seems to be teaching the model what the body looks like again rather than overfitting.

If this all goes well, this new model will encompass the Mooning and Upskirt datasets with this true model of the human body, which should be the best nude model so far.