Frick why did I read this comment at 3am. I actually hurt my back from laughing so hard

ya, I want to do one that turns vag into dick (and more masculine shapes in general), but not giving beards and hairy legs.

Can we make that somehow?

@Janet Something like that would be fantastic. Magical, even. With a lot of LoRAs, if they aren't overcooked you can substitute a male for a female fairly well. I've tried it on fellatio and penetration LoRAs with better results than I expected. But it's pretty variable and LoRA-dependent. It's also a lot harder to prompt out things like breasts. And in some cases, it's possible to even give the character a penis... but often there will be remnants of vulva too, below the testicles, resulting in a sort of male futanari.

As for the hair... For the most part I find that putting "(hairy, body hair, facial hair :1.4)" in my negative to be fairly effective. If that doesn't do it, you can try being more specific: "(hairy, body hair, facial hair, chest hair, leg hair, navel hair, arm hair, back hair, ass hair, pubic hair, armpit hair, beard, mustache, sideburns, goatee, stubble :1.4)" and/or raise the weight. I like a little bit of hair in certain regions, so what I do is add "sparse pubic hair, navel hair" to my prompt and "(hairy, body hair, facial hair :1.3)" to the negative.

But as for magically changing genitals, probably the only way to do that is to train it on images with penises and call them "female". And I still doubt that would quite be enough. But, it's worth trying it if someone is interested in seeing where it goes.

@dita I've had the same though, train over images with dicks but tag them as 'female' or '1girl'.

Maybe a TI worked better here? I have no idea how tokens etc... work but apparently they just shift weights in a certain direction, so if we trained a TI that moved in the direction away from vaginas and toward dicks, I guess "discovering" which tokens move the weights in that direction?

I've tried, by the way, putting vagina Loras into the negative, or using them with negative weight, but not luck. I think someone smarter than me needs to figure this out.

@Janet Yes, it's easy enough to find which vectors (numbers) correspond to the tokens (words/word parts). There's even an extension for Automatic1111 webui that will translate tokens to vectors and let you alter the weights for specific vectors inside a TI. So it would be a easy test for someone who knows how to do it. I merely know that you can do it, not how. 😆

For using a vagina lora in the negative, probably a better idea is to keep it in the positive prompt and give it a negative weight, like <lora:yay_vulvas:-1> or something. My SD is a bit funky at the moment, so I doubt I can try it. If you're up for it though, I'd love to know how it goes.

@dita ooh, I tried it, and it works somewhat. I tried the Asian pussy with a negative weight. At -1 it turned the skin yellow and made the quality go lower, but it did indeed remove the vagina...

That gives me hope. Like, what if we trained a model on low quality yellow jpeg vaginas, and then used it as a negative weight? That could increase image quality and decrease vagina at the same time? Or maybe as you said earlier, better someone smart who could work directly with tokens.

@Janet The reason for the yellowing and the low quality is probably because it was too heavy a weight. In that case, I'd try something like -0.8 to -0.6. There just might be a sweet spot where it will remove the pussy but leave the quality and style alone.

@dita ya, and it might take a few different vagina Loras to see what works.

But wouldn't a TI make more sense in this case? Maybe there are only a couple of tokens that move the weights toward vagina, and these need to be negated?

@Janet Yes, a TI would probably work better. The thing is that LoRAs generally don't apply to the negative prompt, which is why you need to use a negative weight instead of just putting it in the neg. But you can put TIs in the negative. I'll try whipping up a TI soon and see if I can get it to erase the vagina.

@Janet So I found what we need—and then I couldn't remember where to find this thread, so it's taken me a little while to get back to this. It's called a LECO: Low-rank adaptation (LORA) for Erasing Concepts. It installs and works just like a LoRA, but the effect is that instead of adding a concept, it erases one from the base model. So using that, it would be possible to erase the concept vagina, pussy, etc. Now we just need to find someone to train it. 😄 Liquidn2 makes a bunch of them, but I've not seen anyone else upload one. To give you an idea, take a look at this one, which makes it possible to limit fireworks to only ground-type like sparklers—or, if you set it to negative weight, you get a sky filled with them.

@dita ya, exactly. There's actually a huuggingface page that says it'll erase concept, and I've tried it, but it wouldn't let me download. This might be exactly what we need. So many things I'd love to erase from a model. (long hair, vagina, makeup).

@Janet I'm sure there are other like-minded folks out there, and maybe someone among them knows how and can and will make a LECO to do those things. Or, maybe someone will publish a LECO how-to or something.

@dita II hope so... but I feel like we're first get 10 Loras which only do the opposite, lol

@Janet That's about 95 percent likely.

@dita there is one other gender slider, btw, it goes both ways, but also makes men only bearded and wrinkled.

@Janet I'm familiar with Zovya's embedding-based and Ostris's lora-based sliders. Are those the two you mean? Or is there another one now?

@dita yes, actually, I'm not so good at using them. How do you think they compare? Maybe all 3 at once? Maybe Zovya's works better as the negative embed? I try really hard to get rid of girly-hair and girly-eyes.

@Janet Have you seen this? It might be a start, but I'm guessing they trained a traditional LoRA on Photoshopped images instead of creating a LECO.

As for how those compare, yes, I often do use Zovya's embeddings and Ostris's slider LoRA together. Typically I'll do something like Ostris's slider in the positive prompt, often with a weight of between -1 and -0.5, depending on whether I want a bit of androgynous features. Plus, I'll often use the Boyish embedding in the positive as well as DeFeminize in the negative embedding. Many times, I'll also include Zovya's age sliders as well, depending on whether I want to skew the results older or younger.

Going back to your very first reply, wanting something to make more masculine but without facial and body hair, Zovya's Boyish plus DeFeminize are probably the best combination, along with Ostris's gender slider if you want to throw that in to fine tune the results a bit.

@dita ya, basically what I'm doing. I think the text-inversions are best for these actually; would be great to have more of them.

@Janet Probably so, since embeddings/TIs only affect the text encoder, and not the UNet. So LoRAs have the potential to warp and distort and do weird things, but a TI will only tell the AI "use more of this token and less of this one" kind of thing. Kinda like the difference between telling someone "Go to McDonald's and order a side salad instead of fries with your Big Mac" whereas a LoRA could say "Go to McDonald's and order fries, but you'll get deep fried ink pens instead."

@dita so we need a TI for "boyish" and a TI for "manly"

@Janet Those TIs are part of Zovya's gender sliders, as well as the DeFeminize negative. I may have called them by different names than Zovya uses though.

@dita Ya, you're right. Zovya does have that... I just want more of them, lol..

@Janet So like "effeminate old dude", "bro next door", and "hairless metrosexual"? 😄 I could go for that. Well, maybe not the effeminate old dude. 😆