It's my first trained LORA based on KREA 2
V4.1 RELEASED!!
Better prompt adherance. FAR FAR BETTER realistic skins. Face expressions are here. Most things are better..
Things bad:
- Masturbation concept not defined completely
- Sometimes strange artifacts can come like in 2:3 you want a sexual image that requires width then change to 3:2...
- for censored (Asian JV), increase strength by 0.1 and try 1-2 times.
- feedback welcome as always, gonna upload more images.
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
Better prompt adherence, better cum, slick etc...
Overall better, be mindful of strength.. most of the time is 1 but sometimes btw 1-1.5..
This one was not the final checkpoint as the final checkpoints are overfitting. Please check this one. I felt this was the best
FAQ
Comments (17)
May I know how many images you used in your dataset and how much training steps for your V4 version?
When you said expressions are included in v4.1, does it mean you have included/merged filter bypass? I think a lot of people are already using one and would want to know so to avoid stacking another filter bypass which could be bad.
No. I have added a complete dataset with manual captions for expressions. No bypass.
I didn’t add anything other than the dataset itself which btw required a lot of time for manual captioning, still not perfect but manageable
i tested it with filterbypassv3 and projector scale loras without any issues but they're completely unnecessary now compared to KNP v2
First good JAV generation i have seen. Good job. JAVs are the best at this nich. train on Rara Anzai, Aimi Yoshikawa, Rio Hamasaki.
Thank you for feedback and recommendation. Will add those.
v4 works great, just made a super simple workflow for it to replace the one I made for v2 — no longer need projector scale lora!
Version 4 looks great. Sadly bleeds into the face a bit when combining with a character lora.
Could you mention examples please so i can improve upon it thank you...
@19doorsside884 I just used it in combination with a self-trained character lora. from 0.6-1 character bleed starts and gets worse the higher. Its not crazy but there.
I concurred with other commenters that v4 prompt comprehension is definitely better than v3 but unfortunately it has problems with character bleed (likely caused by adding expressions). Unfortunate because character preservation has been this LoRA's biggest strength.
can you give example prompts please? What characters
In my testing i didn't see any character bleeding, hell i even did a private parody
I tried V4 with some of Turbo's built-in names and just V4. The bleeding is definitely there but I am really anal about likeness so it may not bother other people. I didn't see bleeding in the earlier versions.
I've been using your new model, and I absolutely love it—thank you so much! I'm also currently training a LoRA for the Krea2 model, mainly focused on NSFW/adult interactions. However, the models I'm outputting just aren't good enough. Do you have any tips or tricks you'd be willing to share? I'm sincerely looking forward to your reply!!!
Hey! Make sure your images are not too many and they are diverse with whatever lora you are creating. For captioning i would recommend doing manually as VLMs struggles a lot.