This is a multipurpose lora (dora actually, so make sure to use up-to-date tools) for corrupting waifus and dresssing them up in latex.
black f4ux-latexis a trigger word for the lora, it also turns the clothes into black latex. You can even try combining it with clothing loras, there are a few uploaded examples.Training images include e.g.
leotard,unitard,panties,bike shorts,crop top,tube top,thighhighs,thigh boots,pencil dress. It can figure out more clothing too, as mentioned in previous point but these should be most consistent. There is also option of havingbodystockingunderneath the latex and there ispr1ncess cutoutsimilar to one from my previous loras.To help your waifu in joining the dark side, you can put a visor on her (
bwvisor)and put a little device on her crotch with optional cable (crotch plateorcrotch plate with cable).If your waifu has already fallen to evil, consider adding a tiara (
p0d tiara), a crystal on her chest (chest crystal) and a glowing womb tattoo on top of her clothes (w0mb tattoo).Elements from 3. and 4. are reliably colorable, see examples.
Works with husbandos too.
All buzz collected (and probably) will be used for a bounty.
Description
FAQ
Comments (2)
Here is a bit of a question for the new DoRA:
1. Is it prone to overfit and causing the model to forget some of the original contents? Read from the articles that LoRA is less likely to do things well, but also "forgets less"
2. How easy is it to train DoRA compared to a LoRA? Is it usually faster and more efficient? How fast can you iteratively test different configurations and parameters?
3. Are DoRAs more sensitive to dataset captions than LoRAs? If so, are there ways to make it easier to train a DoRA to focus more on certain characteristics? What about AIOs or styles or characterpacks?
4. How about usage, is it slower to generate than a LoRA when stacked with other LoRAs (e.g. characters, art styles, poses, scenes)? Can this be mixed with Turbo or Lightning?
5. What is the future of DoRA and other PEFTs when dealing with DiT models like Sana/PixArt/Flux/AuraFlow/SD3.5/HunYuan?
I was just experimenting with it back then. Well, from my limited knowledge and experience:
1. Dora is supposed to have similar results to finetuning and then extracting a lora. In theory it makes it possible to learn better but transfer to other checkpoints worse. I didn't notice huge difference, especially in the second aspect when it comes to pony models. From what I experience, I could achieve similar fidelity and less overbake when using dora and lowering LR a bit.
2. Just switching a flag in ui. It's slower than training loras with same parameters from my experience.
3. Sorry, no idea.
4. It shouldn't stack worse, didn't notice it anyway and I stacked my doras with other loras a lot. No idea about turbo/lightning, never really used those.
5. No idea.