This LoRa will allow you to use your favorite character with it without altering its face.
Makes your character surrounded by dicks. Not meant to have dick insertions, I found that tricky with the POV Blowjob LoRa, so keeping it for a later version.
V1.1 : I've added 5 more epochs to it and fixed the training. To be honest V1 was really good for a mistake version, but this one is what I intended. Does not alter the face and allows more flexibility with the expression, depending on your character Lora.
V1.0 : a bit face-altering because I've trained it at the same time than POV Missionary so I've done the same mistake in my workflow (included the faces). It still works very well.
Trigger word : Blowbang
Important words : She is surrounded by penises, She is holding penises with her hands, A penis is touching her face, A penis is touching her hair, Viewed from above, She is kneeling
Description
FAQ
Comments (21)
Did you use civitai trainer?
No. Kohya scripts
Again, great job.
This LoRA seems to be quite a bit stronger than your previous work (ahegao, cum on face v1). Perhaps this is by consequence of being trained on dedistilled? I found that I had to lower strength to 0.6-0.7 for my character's likeness to not be affected. I haven't tested extensively so this could just be me.
You'd know a lot better than me, but would training without poisoned words like "penis" not result in faster convergence?
Thanks again for sharing.
V1.1 did not affect the character face in my tests but I appreciate your feedback. I hope it's ok since it's my goal not to alter characters :-) Regarding the poisoning I never though about it, could be a good point. For me flux doesn't know what's a penis at all, so I didn't mind. You probably mean I should dick instead ? I'm not sure it's important but that's worth double checking
@valentinkognito365 My understanding is that certain tokens were actually poisoned as opposed to not being trained. So in an effort to not have Flux generate images of penises, for example, they would have tagged many random innocuous things as "penis". As a result, the concept is completely misrepresented. It may appear like Flux doesn't know what a penis is but it's way worse: it thinks it knows what it is. It's just confidently incorrect, because it thinks a tree is a penis, and a kettle is a penis, and a vase is a penis, etc. So when you're training Flux to generate penises around someone, it's mixing loads of concepts from its poisoned past when it was gas lit about what penises are.
So my recommendation for something to try, is not using any NSFW terminology in training. Invent tokens that Flux would never have heard of. And I wouldn't be surprised if tokens like "p3n1s" have been poisoned. So I'd try something the developers would never have anticipated, like "d1ngd0ng", just to be absolutely certain. I think running training with the exact same config and images but making these changes for penis, tits, etc. in the caption files could yield very different results. If I'm wrong, the output will be the same.
Let me know what you think :-)
@272727 it's interesting and doesn't cost much to try. I wonder if people have verified this. Lol you made me laugh with d1ngd0ng. Do you know the "ding dong song" from Gunther ? The chorus goes "Owww... you touch my tralala... Owww... my ding ding dong". Would make funny prompts : blowbang, tralalas touching her face, dingdingdongs on her hair, etc :-D hilarious
lol that even gives me the idea of training Gunther so I can have his face on males fucking girls
@valentinkognito365 PLEASE do this hahaha. I love the ding dong song. When I was typing the poison theory above I initially typed tr4l4l4 before I changed it to d1ngd0ng. Can't wait to get prompting with this - I start growing my moustache today.
@272727 yeah the more i think about it the more i think i will actually do it. We could generate GTA-6 pictures of him
@272727 I gathered 20+ pictures of him, and good quality for most. This is happening.
@valentinkognito365 haha amazing news. I was thinking about it and the prospect of trying to find decent photos of 2004-era Gunther. Well done
@272727 it was surprisingly easy. I will release it in about 2h+ time if everything goes according to plan. You can already think about your prompts lol
Quite interesting though I only get hands around her... which checkpoint should I use? Thanks
The suggestion is dedistilled but you shouldn't be getting only hands with the original flux dev checkpoint. Check that your prompt is in line with the examples and try adjusting parameters in your workflow to ensure that the LoRA's influence is strong enough.
I've found out you might have to decrease the strength of the character to avoid having closeups on the face, or prompt more towards above view / full body view. The LoRA works no problem with distilled, it's just it was trained on dedistilled and everything is better on dedistilled anyway
you managed to crack the NSFW LORAs. Whta is your secret master?
Motivation :-)
@valentinkognito365 I appreciate. But seriously. You just have pics captioned in joy cpation, you feed them to de destilled FLux and than apply this a LoRA to the normal FLux?
@barlogspam176 Flux Devs love dick so its not censored.
awesome lora! thanks!. it seems they never shut their mouth on my testing tho.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.