Just a rushed Vagina lora, only 3000 steps which took 4 hours to train on my RTX 4090. Training data was focused on vagina images; 32 images. All the data was closeup shots of vagina, while the lora performs exceptionally well in close-up shots, it performs poorly in full body shots because there was no full body pussy training data. With a larger data set of both close up and full body I think we can train something amazing. This lora performs well at the specific poses that were in the training data, but I missed a lot of poses, so it will perform poorly for those poses. Someone with a far larger data set will likely come along and fill in the gaps.
If you're seeing a lot of stretch marks and tanned skin its because the training data was of it and Ideogram seems to take to the data extremely literally. In hindsight I should have mixed the data set with more pale and other skin tones.
I'm not an expert lora trainer, just screwing around. Ideogram Loras seem to be easy to train, it adopts data extremely quickly. Far easier to train than Z-image IMO.
Limited testing done. Weights that work for me are
0.9 COND and 0.5 UNCOND Lora
0.7 COND and 0.6 UNCOND Lora
0.6 COND and 0.5 UNCOND lora works well too, probably the lower the better as to not overcook the image.
Euler = speed
Res_2s= quality
Do not make both loras the same weight, like 0.6/0.6, for some reason the model starts to hallucinate badly when that happens.
Ive included the AI ToolKit log file for anyone interested in how the training was setup.
I think the reason why my vagina did not properly converge on full body shots is like due to me using adamw8bit and linear, which is a far more tame training method. My next run I will use the more experimental aggressive optimizers.
I will not be able to do massive 20k+ step training runs, my runs usually end around 3-5k because im an unemployed gooner that just loves AI but has to budget my power bill. I just happen to own a 4090 because I went broke to get one, totally worth it though.
Description
Early test concept of an innie vagina
FAQ
Comments (9)
Great Lora, was hoping for a one since I couldn't prompt Ideogram 4 to generate one. Will there be a newer version with less tan and stretch marks?
I dunno, I only do these things on a whim. There's no rhyme or reason to when I decide to train a LoRA, unless someone beats me to it. There are a few dedicated vagina lora trainers here that pop up once in a while that are much more talented and experienced than me. As of right now, I'm not really too motivated to start a new run. Training a lora is the easy part, the hard part and most time consuming part is curating the data and captioning them(or spending tons of hours unfucking the bad captions AI made).
Good first version, but it changes the girl too much. Maybe try making your dataset more varied. This was making slender ladies in butch ass bitches.
Butch as in ugly? im getting some very pretty girls in my outputs.
@fox23vang226 Yes. Butch as in quite manly. Realism isn't hurt, maybe even improved, but it's manly realism which is not idea. I think the lora learned far more than pussies.
@gjsrgfnmasosseonkt179 I dont see anything manly about the sample images at all. Are you Asian and used to seeing extremely skinny and soft frail bodies? not used to seeing thicker European types?. I kinda like my women with an athletic build.
@fox23vang226 I mean, we can call it athletic build, sure, I think it's a bit much. And yes, soft Asian women are objectively awesome. BUT... This is a pussy lora. It shouldn't bleed over into body type at all. You should have so many different types of innie pussy pictures in your training data that it doesn't pick up anything else.
It adds that brown shading around ever girls anus. It's great that there are ass holes, but they don't all look the same like that...
The dataset was very small, of only one girl. If I update it in the future I will add more variety.