It's my first trained LORA based on KREA 2
V4:
Will be released in 1-2 days, one of my friends gave me access to his h100.. This time will work on better prompt adherence and get few issues resolved.
If getting weaker results then use strength of 1.5 or 2.
V3 Released!!!:
strength: 1-1.5
for jav use 1.2-1.5 so mosaic doesn't come.
Added: Asians, Blacks, realistic skins etc... (use to know better)
Recommendation use: https://huggingface.co/Beinsezii/Krea-2-Turbo-Projector-Scale-LoRA-Diffusers for better control with 0.01 (but it takes hit on quality just a little)
For better prompt adherence, increase strength a little of the second lora...
also, don't just write a thick cock between or somewhere.. write like a man then his cock so it understands.
Prompt adherence can be a hassle please use the above lora (0.01 weight) and next time I will try to make it better. V4 will take some time since gathering dataset, 200+images with 5k steps and training took a while and lots of cash from me, trained with RTX 6000 Pro.. Please give feedback since I am learning...
V3:
Will be out within 1-2 days, gotta arrange those gpus. Also, i got a massive uhd dataset i found somewhere, so this version should be most realistic and the best version.
TO BE FIXED:
- Mosaic Issues sometimes
- Asians to be added
- Blacks enhanced
- Better skin (sometimes skin looks plasticy)
V2 OUT:
Recommended:
- Strength: 1-1.5
- Do NOT USE conditioning node as it seems to be degrading the quality with the lora.
New Features:
Better SKIN, no need for KNP trigger TAG.
Added:
Detailed Pussy
Detailed Assholes
Different Poses
Different undressing styles
Male Cock
Sex (Basic)
Cowboy etc...
More things to come especially sex styles and art loras soon but :( it took 5 dollars to make this...
Please give Feedback.
Description
FAQ
Comments (73)
Great! Thanks.
No Issue. More versions will come soon and will be better
Comparing the images to the prompt is a good laugh lol
:)
Not working for me. What strength are you using?
Try 1.5
Quite good, help not only with NSFW, but also with overall poses.
Will there be anything to make sex poses easier?
yes, it's being trained
Does this work for Krea 2 Turbo version?
yes
the first lora and is NSFW, based
hey man, sorry but working on SFW lora too. Art style... Any recommendations?
Why you think AI was born?
@19doorsside884 i didn't say it in a bad way, i congratulate you for it :)
train huge breasts, like exagereted, and much more xd
thanks man, finally a real solution to do nsfw
No problem man. Check the V2 and please give feedback
This is my Opus moment. WTF KREA cooked and your lora is great!
Thanks man your words will keep me running
this lora works so much better than I expected
Thanks
Insanely good
the model is fully uncensored. just look search for the lora on reddit that uncensored to model.
Great job, Loras! If possible, please include some datasets of Asian ethnic groups! Thank you very much!
Yes will do in next lora
@19doorsside884 We are extremely excited! Thank you for your efforts. Thank you very much! I would like to ask if you plan to train a separate branch called "lorass" using the Asian population training set, or if you intend to add more Asian population datasets to the existing "loras"?
Thanks for the lora, but please rename your files before you upload them :D
KNP_V2_copy_copy
I am not going to remember what this is.
Ypu're saying we shouldn't use this Lora along with the ComfyUI-ConditioningKrea2Rebalance node, right?
that's exactly what they are saying.
Yes because that node gives plastic like skin and makes the characters look strange
that node is useless, it tweak the layer of the model aka affect everything output, not just censorship
working great within WAN2GP with KREA 2
it is working
Damn it hasn't been 24 hours yet . . . these people smh
what dataset are you using?
Made my own data source from many sources like cornhub etc…
Just one useful comment: it works better AFTER this additional mini-lora, mentioned in reddit post - https://huggingface.co/Beinsezii/Krea-2-Turbo-Projector-Scale-LoRA-Diffusers (tested that mini-lora with Strength 0.5 and such combination removes all "spare parts"+ better listens prompt)
Whit what strenght?
Thanks, that seems to help.
@Addictedd I am currently quite satisfied with 1st mini-lora (by Beinsezii) strength 0.5 and 2nd lora (main) S1. But it's up to you to test and compare alternative strengths ;)
this lora gives me black image at 0.4 str
@tugrikyan ye its good, i tried
Works pretty well with characters Loras, thanks!!
interested in your training parameters, if you're willing to share? Ostris AI toolkit?
Did you use censored JAV photos in your data set because I'm getting an annoying amount of mosaics in my outputs whenever the subject looks Asian
So sometimes it seems censorship issue is there in this lora i guess some bad data. Will fix in next Version. Thanks for feedback also gonna add Asian lora
oh.
ok.
OK.
OOKKKK!!!
WHAT?!?!?
EEEYEEEEEAAAAAHHHHH!!!!!??!!!!!!!
Is it bad? 😢
Great work! This lora definitely needs more hentai data. Maybe it just needs more data. To be honest, we need a full finetune on a huge dataset, but it's just too much to ask rn :).
Also it kinda struggles with the masturbation concept.
yeah masturbation was not in dataset... will add in the next one
Great job and so early too ! And to think i just had caught up with my ZIT setup... Crazy how fast you guys have trained these early versions. I that because the team gave insights as to which values to use ? It took so long with ZIT even after the ZBase release (I can't train stuff so i'm particularly candid here)
I think you mostly trained on pale skinned people, right? It gives some darker skinned characters white skin like they have vitiligo on their breasts and ass.
yes will add darker soon
This lora has significantly increased the generation time. Is it normal or am I doing something wrong?
Seconded (on SwarmUI), I go from 20 seconds to literal minutes per image :/
(krea2_turbo_fp8). I remove LoRA, & it instantly goes back to normal.
PS - Might be issue with SwarmUI
Update - Tried another LoRA, it does the exact same thing. I'm on the latest update (which said in update notes that it added support for Krea 2 LoRAs).
Seems an update for SwarmUI/Comyfi (possibly related to custom extensions) is making it more sensitive to going over the VRAM limit.
https://huggingface.co/Comfy-Org/Krea-2/tree/main/diffusion_models krea2_turbo_nvfp4.safetensors 7.67 GB
@ShadowCell I'm using it on ComfyUI
@Shigmund SwarmUI uses a ComfyUI back-end.
I did a fresh install of SwarmUI, & it fixed it properly btw.
Might be related to ip-adapter-faceid, or another extension add-on causing an unintended bug.
Results are random.
can anyone please provide working workflow file?
Works better than i expected.
Works great...no issues at .8, euler simple
Thanks
使用该lora有概率出现马赛克
When generating Asian characters, it seems this LoRA tends to produce images with mosaics, just like JAV. Is it because the training data for Asian characters was censored?
no data was censored, it's a strange behavior i found. Will not happen in the next version. Please stay tuned ;)
It is about the model itself. If you use Comfy, you can test with the Conditioning Krea2 Rebalance node without lora and increase the bypass multiplier.
This LoRA works better that the condition rebalance patch, thanks for sharing =D
how did you caption your training images?
I'd like to know this too, please. :-)