This model was trained on the following concepts:
• Crotch bulge
• Crotch bulge with exposed testicles
• Crotch bulge with shaft outline
59 training images @ 10,000 steps were used.
This is the first concept lora I've created and it is imperfect. The "crotch bulge" concept is by far the easiest to get usable results with. The "exposed testicles" and "shaft outline" concepts may require multiple generations to perfect. To speed up this process, once you have a bulge with the general shape you want, you can use variation seed + strength to hone in on a passable result.
No masculine/male characters were used in the training data, nor have I tested it on masculine/male characters, nor do I have any interest in doing so.
I found that SwarmUI's <segment> prompt syntax was helpful for cleaning up bulge image outputs and giving a more realistic/smoother result, but I suspect any Adetailer-like cleanup step would also help.
This model also works very well with inpainting/editing workflows
Additional keywords such as "big crotch bulge" or "round crotch bulge" can also be used to affect the image output. The model wasn't trained on these, but Klein seems to natively have some understanding of them.
If you want to help me perfect this concept by paying for runpod costs, feel free to leave a tip:
https://buymeacoffee.com/daviddickloras
Description
FAQ
Comments (5)
Thanks. Are you considering doing a Z-Image version? I've tried using a similar lora by another creator but it didn't work nearly as well as this one does.
I've used Zimage a handful of times and wasn't as impressed with it as I was with Klein. But I've also never tried lora training with it. I would be interested in seeing how it does with preserving multiple concepts like exposed testicles/shaft outline. I'd like to add some additional concepts as well, like transparent/mesh underwear with dick visible underneath. So, perhaps I'll give it a shot at some point
I can give you a few pointers on things that I found improved Z-image Turbo, if you like. For one, there are loads of great sliders that can give you granular control at seemingly no cost e.g. you can stack as many as you like and they sit happily together unlike most loras. If you want any other of my recommendations, I'm happy to share. Flux Klein remains the king for quick adjustments to images but ZIT seems to be the best for "realism" in my opinion and gives a bit more variety of faces by default. Anyway, this lora works great with Klein - could be amazing with ZIT, I reckon. Thanks again and I respect your decision either way.
@DavidDick Think you'd make one for Krea2?
@Tsarcasm I have one for Krea 2 ;) I'm trying to tweak it a bit but I'll release it soon