This is a textual inversion version of my original Dreambooth for women in their underwear. It offers greater flexibility to be used with many different models based off of SD 1.5 though if you want quality at the expense of flexibility and variety, you should also try the Dreambooth.
I created this textual inversion specifically due to the ubiquity of models for nudes but there were very few for sexy clothed women. As the name implies, it's designed to be used with models for nudes (like F222.)
It's trained on a similar dataset as the Dreambooth using the original SD 1.5 model and on 16 vectors. I've had great results using this on F222, Hassan Blend, Dreamlike Photo Real, and even the Ally's Mix (for anime pics.)
I've moved the recommended settings to the respective model pages since they're slightly different for each.
For more tips on how to create beautiful women using SD (not just with this model), I've written a guide here.
Description
The trigger word is: shirtp5
NOTE: every version has a slightly different trigger word. This one was retrained on a much larger dataset of about 100 images with with greater variety.
(Yes, the model was originally designed for women wearing sexy shirts but I had gotten better results with underwear so I pivoted towards that.)
Most of my testing of this model revolves around the following settings so this would be a good place to start:
Recommended Settings
Model: SD 1.5, Hassan Blend, or F222
VAE: vae-ft-mse-840000-ema-pruned.ckpt
Hypernetwork: None
Starter Prompt: woman, perfect face, standing, t-shirt, panties, feet, shirtp5, bedroom
(Note, at a bare minimum, you'll want to mention the following 2: woman, shirtp5)
Negative prompt: small eyes, eyes closed, long legs
(It seems the images I trained the model on had people with smaller than average eyes but putting these in the negative prompt seems to help a lot. And depending on your dimensions, the legs may sometimes come out unnaturally long so include "long legs" in your negative prompt if that's the case.)
Height: 1024
Width: 512
Sampler: Euler A
Steps: 20
CFG: 10 (though anything above 7 seems to work fine)
Restore Faces: On (makes a huge difference)
Highres Fix: On, Denoising = 0.5
FAQ
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.