Sign In

Shy Undies - Waistbands/Sagging Pants

124
827
4.1k
54
Updated: Sep 27, 2024
clothingfemalegay512x512outfitman
Verified:
SafeTensor
Type
LoRA
Stats
98
42
Reviews
Published
Sep 25, 2024
Base Model
Pony
Training
Epochs: 10
Usage Tips
Strength: 1
Trigger Words
wband
waistband
Hash
AutoV2
77F7B02E3B

Plenty of checkpoints render people going commando, with low-rise pants stopping just shy of the dangly bits. With this LoRA, you should be able to put an extra layer on your subjects and let them show off those fashionable (overpriced) undies.

The LoRA has been trained on roughly 200 images and plays well with a number of semi-realistic and realistic checkpoints.

Want to provide me with direct feedback and/or share what you've made outside of Civit? You can find me and other like-minded folk over at the ctrl_alt_diffuse Discord.

Caveats:

  • Waistband branding, logos, are hilariously garbled. (Improved in v2)

  • Some waistbands can show up extra thin, or extra thick. I haven't sorted out how to address that yet, so YMMV.

  • While it was trained mostly with male references, it works well on women too! Sometimes the checkpoint will add straps instead of a waistband, but that can be a good thing.

  • The LoRA has the accidental benefit of adding unique asymmetry to non-modern concepts, such as fantasy or sci-fi. It typically ends up rendering angled/lopsided wraps, belts, or similar waist adornments. Definitely a happy accident.

Updates for v3: My testing has had greatly improved interactions with a number of checkpoints with this new version, as I have refined the training images and tags significantly. With v3, you should start with a weight of 0.6, and make sure you include some kind of outerwear in your prompt, e.g. pants, sweatpants, trackies, jeans, etc., as it will likely only generate underwear.

Updates for v4 (Pony): Even in Pony, for all of its strengths, it's sometimes difficult to get the waistbands to show. So, as I am wont to do, I have created a new version of this LoRA to bring them back into view.

In this version:

  • Same dataset as v3, upscaled to 1024x1024

  • Depending on base model, you can go as low as 0.5 (maybe even lower!) and get the results you want without compromising quality.

  • Combining keywords wband with waistband can help in some cases and can hinder in others. YMMV.