I was not happy with how checkpoints handle pantyhose when the subject is not wearing any bottoms. I trained this LoRA on various pictures involving women wearing nothing but pantyhose, wearing tops but no bottoms, or wearing a dress and pulling the dress up.For best results:Prompts should start with some variation of the phrase "woman standing in pantyhose". Most of my training data involved the phrase "young woman" specifically, as models in pantyhose seems to invoke older women in stable diffusion checkpoints. "Naked woman" or "Bottomless woman" are also very common phrases in the training data. Training data always includes a color before the word "pantyhose", with "nude pantyhose" used for colorless pantyhose. Opaque pantyhose and shiny pantyhose are also possible options.All LoRA training data used 768x1344 image ratio. I imagine it would work with other ratios, but this is worth keeping in mind.This LoRA is meant for use with realistic checkpoints. I tried it with some anime checkpoints and it didn't really make it look any better.Lower CFG scale (around 4-4.5) seems to give a more realistic look, but it can appear a bit uncanny. Higher CFG scale (around 9-9.5) seems to give a more glamorous look, but less realistic (in a photoshopped way). I like to start around 7.5 CFG scale, as for these sorts of images, I think the glamorous style is more suitable. I trained this model very slowly, as a normal training speed did not give the results that I wanted. Therefore this model will improve images in a subtle way. I plan on running it further and adding a less subtle version of this soon, if the results are worth uploading.

Bottomless Pantyhose

I trained v9 with the same dataset but with Prodigy optimizer as opposed to AdamW, as I've heard it gives better results. The upside I've noticed is that yes, the results look noticeably better. However, much of my dataset involves pictures of women with their face cropped out of the shot. I tagged these photos with the phrase "head out of shot". However even when including "head out of shot" in the negatives, it will still often times produce a result with the head being cropped out of the shot. If this is an impediment to the result that you want, you will either need to lower the multiplier for the LoRA or go back to v7.

Files with SHA256: a11b8729b2f9bb1880b6a767a77e1b445590e0376d3dc8108e3e30c6c3416455