Tied Standing

While collecting the tied spreadeagle dataset, I also happened upon lots of other BDSM. This concept was easy enough to cut and caption so I chose this for my latest concept.

The dataset was, unfortunately, rather limited, as one production house was responsible for the vast majority of the high quality videos featuring this pose. This meant that the backgrounds were very redundant. Another long conversation with ChatGPT has hopefully expanded my skills with captioning to avoid inflexibility in this instance; lessons I applied to version One.

Versions

One

All around a very flexible model given the limited dataset. The new technique I tried seems to have worked very well. You still need to prompt properly to get what you want, without being afraid of over describing your desired outcome, but this LORA doesn't get in your way.

Trained on 12 288x288x50/2sec videos at 5e-5 learning rate. Minimal captions with strategic noise, variance and misdirection tokens.

I've written a training guide, so check it out if you want to get into it!

Disclaimer

You can run, but

you

can't