A early test model for a hogtie/barefoot Lora trained on Realistic Vision. It takes quite a bit of fiddling to get a decent image in it's current state. Uploaded just to show progress.
Edit:
I'd suggest trying to start with one word, i.e. hogtied, prone, kneeling, etc, and then add a word or two to try and add what you want, to see what works.
Description
Trained on a larger image set using Realistic V5.1. Still takes a fair amount of playing around with to get good images, but less so than 0.1. I used strength around 0.5-0.7 depending on how much trouble it was having getting the position right. steps 40-60 generally worked. Usually 512x512, but sometimes 600x600 also worked. I mostly tested it with Euler a and using the latent upscaler at 1.8-2x resolution, with a denoising of 2.5-3.5.
If the image doesn't look quite right, slight changes to the strength, steps, the denoising could change it enough to potentially fix it.
Important words for prompt are hogtied, prone, kneeling, side view, front view, rear view. I don't have enough images for the top of the foot in some angles, so it can't make it look right, or it just tries to put of the sole of the foot on top. For the hogtied prompt, you can also add arms tied, legs tied, toe tied, hair tied, and rope chest harness.