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

This will likely be the final version. I improved the tags in the training images to remove some typos, and changed the phrase "facing the camera" to "facing the viewer" as the former would result i the model holding a camera sometimes.

Files with SHA256: e4643057b2038216f903870ac586b25439ac5b073c98f999826a1993ec670cac