Krea 2 didn't understand much about this but it didn't take much to open up its abilities.
Version 1 of this for Krea 2 is limited to mostly hanging, from mouth, hand etc. The "drinking from straw" works quite well, though cherry picking may be required in order to get the physics right since it likes to make it just float in space.
If there's any interest I'll make a more versatile one, what would you want from this type of concept/object?
Version 2 of this LoRA for Krea 2 is quite a bit more versatile but trained on a different dataset, and larger set as well as rank (120/32) amounting to 3,000 steps. I have another one baking for an additional 3,000 steps to see what happens but I don't expect much from it, though it may give me some guidance for an improved version.
I haven't explored its full capabilities yet but the dataset included everything that was requested, or offered as an interest, in the comments up to this date, Aug 2, 2026, but I can't promise that it can produce that content.
Version 3 of this LoRA was not a fun experience but I learned a few things so not a waste of time. I had to settle on getting most of the concepts working but "condom emptying" may not preset very well, which I'm making a separate LoRA for, and I got the occasional hit for prompt changes as simple as the color of a garment. However, this version eliminates redundant images and those with the worst quality, resulting in a more vibrantly colored potential and includes a better presentation of ...
. condom hanging over something
. condom across the eyes
. more interesting, and visually coherent, condom belt type things
. better color separation and vibrancy
Missing is condom emptying but I baked in a term "upside down condom" that can sometimes simulate the effect, though it's not often complete. Again, I'll have a separate LoRA for that if only to discover how difficult it is with that sole concept.
Description
Version 2 of this LoRA for Krea 2 is quite a bit more versatile but trained on a different dataset, and larger set as well as rank (120/32) amounting to 3,000 steps. I have another one baking for an additional 3,000 steps to see what happens but I don't expect much from it, though it may give me some guidance for an improved version.
I haven't explored its full capabilities yet but the dataset included everything that was requested, or offered as an interest, in the comments up to this date, Aug 2, 2026, but I can't promise that it can produce that content.
FAQ
Comments (5)
Okay, tested v2. I didn't want to post on the other comment area because in case you miss it. I usually test loras extensively so when I comment, I make sure I don't speak out of my ass about it. haha.
Anyway, v2 works really well but unfortunately I couldn't get the same results as you did. I tested everything you did in the prompt. I went close to it but not exact. For example, the condom that dangles on the guy's penis seems much better on your image than mine. It's defined where's mine looks like it's integrated with the guys penis and looks more like cum than condoms. What model did you use?
Overall, I think this one does exact what it promises but for some reason (maybe prompt issues?), I can't generate the exact ones.
I might be able to find the issue if I knew which image since I save every example.
But maybe this helps, I use a resolution of 1056x1344 in order to fit nicely with CivitAI's model display. Additionally I have 8 ksamplers running different schedulers and samplers, I suspect this might be the area to test.
I'm currently baking a new LoRA from scratch, I wasn't happy with version 3, which I never posted, and maybe this will be more versatile.
Thank you. Appreciate the hard work and dedications.
@jth3gravedigger625
I think I found the image, is it with the punk lady? Anyway, that one produced the result you're talking about, if I understand you correctly, consistently with these options...
sampler: euler
schedular: beta, sgm_uniform, simple, normal
sampler: er_sde
schedular: simple
I put the LoRA I use in the prompts, commented out, this is one of them along with the weight...
// snofs_krea_v1 : 0.63.
And of course the main one...
// used_condom-save-2999 : 0.90
3 out of 8 didn't hit, on some runs, so I wouldn't call this cherry picking or difficult, although I do spend up to an hour composing a prompt and picking an image that I like best for each of the 4 initial display images but the one you mentioned hits at least 5 out of 8 times using euler/normal and euler/simple.
And thanks for your feedback, it's more appreciated than you can imagine.
@Tarterbox Thank you. I will give this a try.