Please read the version details for updates, thank you.
"I like girls that are like little parade floats" - Anonymous
Produces girls with a body that looks like a hyper bastardization of the venus bodytype, enlarging it and thickening it to the extreme, while taking care of keeping the overall hourglass shape. While the girls it creates are beyond a doubt overweight, it attempts to avoid excessive fat rolls, giving the impression that the midsection is larger as the foundation for the much larger top and bottom halves of the body, rather than being conventionally obese.
The head is smaller as a shorthand for making the generated girls look truly massive. Because of this, the faces can look shoddy unless the angle allows a closer view of it (such as with high angular shots, cowboy shots), in which case it is recommended to lower the weight when upscaling.
In general, this model can also help you if you enjoy gigantic breasts, asses AND bellies, but you struggle generating all three in a single picture.
This model was created using Prodigy, a dataset built on roughly 95% AI pictures generated with earlier versions, and most notably, no tagging whatsoever, as an experiment. Therefore no specific keyword is needed, only guiding it towards the desired shape through thickness tags. I listed a handful of those as the trigger words, even though the model hasn't been trained on them, for your convenience.
Muscular girls with this body shape (strongfats) can also be made, my go-to method for those besides adding muscle tags is to also include "abs" in negatives. Be advised that muscle and tone is correlated with slimness so they might compromise upper size.
Also mind that backgrounds fight for space with the character so you might find it easier to create a bigger character in an (unprompted for) simple background than on an specific location.
Description
First try training for a v-pred checkpoint - in particular, NoobAI V-Pred 1.0.
The dataset was extensively revamped. More than half of the examples were cut due to low quality (mostly low resolution 1.5/early PonyXL stuff). Most of the surviving pictures were upscaled and/or overhauled, and some higher quality examples were added. This should result in a softer, cleaner look with less umprompted cellulite/folds on the legs, as well as fewer instances of linea-nigra and shiny "flab" abs.
In a vacuum, it works well at 0.8~1.0, with a "base style" still reminiscent of Orange Mix. However, It is very sensitive to style and characters and they will limit upper size. I figure keeping things on model is a v-pred model advantage, but that isn't particularly helpful here. With characters, additional thickness tags work well enough, but with styles, you may need to lower the style's strength or raise it on this lora. However, around ~1.3 strength and above it muddies colors and causes anatomy errors.
In any case, thickness tags such as those suggested will help and are preferred to raising this lora's strength excessively.
For what's worth, it appears to recognize the class token "jmbvns" a little better than the non-vpred versions (better than nothing, that is) so you might want to throw that in.
I plan on training non-vpred Illustrious and Pony versions with the revamped dataset soon-ish.
FAQ
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.