SDXL
v2 is an adjuster/slider model that comes in a SDXL and SD 1.5 model variant. It's trained on front, below and from behind images. It was trained before pony and illustrious models existed but seems to work on many mixes if the weight is set high enough (more than 2 on some models)
Krea2
v8a is trained on krea2 raw. I recommend using krea2 turbo for inference. Other nsfw models can be used but aren't required.
Flux.2 Klein 9B
v7a is trained on Flux2-Klein-9B. It was trained as an image-to-image model but also seems to work for text-to-image. The main keyword used for training is "fix her crotch" but it is not required for it to work. To simplify training, unlike v6, this model was just trained with the slit pussy type. I may release other variants in the future.
v7b is a Flux2-Klein-9B img2img model similar to v7a, but trained on an innie pussy with a protruding clitoral hood.
Z-Image
v6 is trained on zimage (and also works with turbo). It's a multi-concept lora. Using "pussy slit" or "slit" will create an innie. Other concepts that seem to work at least some of the time in testing include "clitoral hood", "spreading own vagina with fingers, one hand", "spreading own pussy with both hands", "penis", and "vaginal sex"
v3 is trained for z-image turbo. It's an adjuster/slider model with no keyword needed.
Qwen
v4 is trained for qwen-image-edit-2511. It should handle input images in most styles including anime or furry content. it should also do a better job of drawing buttholes than previous models and also supports images with anal penetration (vaginal penetration not trained). The main activation keywords to use are either "adjust her pussy", "adjust her pussy and anus", or "adjust her pussy, anal"
Chroma
v5 is trained for Chroma. It works best when the keyword "cleft of venus" is used.
Stable Diffusion 1.5
v1 is an SD 1.5 adjuster/slider model. It's trained only on low angle images from the front.
v2 is trained on images from more angles (note there are 2 different v2 models. An SD 1.5 version and a SDXL version)
Description
Another Klein I2I LoRA. Protruding clitoral hood edition
FAQ
Comments (28)
⭐⭐⭐⭐⭐
version b is 🔥🤩
If I use this, I get grainy results with Qwen. doesnt matter if I use qwen-image or qwen-image-edit.
I noticed that with qwen when testing with the 4 step lightning lora. Try running without lightning (remove the lightning lora, change steps to 40, change cfg to 4) and the graininess should dissapear.
For editing, I'd recommend trying the klein model as it runs many times faster than qwen and is more likely to produce a decent result on the first attempt.
@m99 I had grainy results even without this lora, sometimes. So I concluded that this lora was not the culprit. I determined that the Resize and Pad Image node worked to remove all grain. I used this node to bring the photo to a resolution that Qwen is optimal for.
Both versions alter the face and style of reference image on flux klein 9b base
It sounds like you might be using the version 7 loras in a T2I (text to image) workflow. They were trained for I2I (image-to-image) workflows and in my testing, don't affect the face/style at all when used in that way. If using klein for T2I, you may get better results by first generating an image without this lora then editing the image with the lora.
The fact version 7 works at all in T2I is a coincidence. Unlike other versions, v7 (and v4) weren't trained to not alter style in T2I as they weren't even trained with T2I in mind to begin with.
Also, while the loras were trained on base, I'd recommend using the distil (non-base) version for editing as it is much faster without any noticable drop in quality (under 2s per edit on a 5090 with the nvfp4 quantization of klein)
@m99 Thank you. I started using "flux-2-klein-9b-RD_AIO" that includes vae and the encoder, and it worked well with that one. It was a mistake on my workflow with the flow of my nodes, it is I2I with multiple reference images so my mistake was also affecting other results, other inputs and loras. I was using "pornmaster Flux2_Klein base V3 fp8". I am on a 3090 I still don't understand the "distill" thing, I may use the bf16 for final version of an image... the RD AIO is heavier, but is fast without turbo lora, but by my few tests it censored the pussy frequently...
@javierflores913 On the topic of the "distill" thing, Klein 9b is a model by black forrest labs that was released in 2 versions:
- FLUX.2-klein-base-9B -- the base model that uses cfg and 20-50 steps. This is the model I assumed you were using based on your original post.
- FLUX.2-klein-9B -- a 4 step "distill" that can generate images without cfg in 4 steps.
That's a seperate thing to quantization, which can improve model speeds on newer cards by processing smaller quants faster (fp8 on ada/blackwell and fp4 on blackwell). AFAIK, on an ampere card like the 3090, you won't see performance gains past bf16 unless you're maxxing out vram (everything slows to a crawl when the GPU has to start using system memory.)
most important lora for every model! using it with flux klein 9b now. works very well
most important that all women in your images have this practically non-existent style of vagina? ya'll have issues, kiddos
@Narz what's the benefit of generating something with AI instead of just browsing the web? Right, you get the stuff which is rare to find, like that style of vagina, or a particular shape of breasts, or a princess who gets fucked by a appletree while driving a race car on the moon, etc. Why should I generate 1:1 the stuff I find on thousands of website in a better version than my AI slop?
also, often generated vaginas are kind of messy because the concept is not clearly understood by the model / lora and it generates random lines and shapes. So it helps to make a cleaner to read shape.
also .... dude, we are here for porn .... created by AI ..... I'd thought that this here is the last place someone starts picking about other peoples fetishes ;P
Amazing lora! Any plans on making a flux1 version?
Since Ideogram 4 is getting its first loras now , this lora would be awesome to have too! You made so many version of this lora. I hope Ideogram 4 will be one of them soon =)
This is the single best NSFW Lora possibly ever created. v6 for ZIT works by itself and you can take the strength up to 2 or more at a certain point it just does the same. I also found it doesn't distort images in any meaningful way, which is really awesome since you can just keep it on all the time as opposed to switch on off depending on what image you're doing. Thanks! The only possible improvement would be to cover even more poses and make sure the p isn't visible when it shouldn't be in some overhead and such.
EXCELLENT lora for Flux2 Klein 9B, works great!!!!
Is Krea2 the next version?
Would be awesome
Definitely need a Krea 2 version, as all the current LoRA's, and even the base K2 model, only create body-horror roast beef sandwiches. All of that wrinkly dark meat hanging down needs to be tucked right back in along with some plumping of the outer majora lips via a proper m99 labiaplasty LoRA. Krea is almost perfect right out of the box... but it absolutely needs some help in this area. Risqué artworks of woman looks so much better when all of the lines are clean and rounded.
Thank you! Pussy and anal for 2511 are just wonderful 🩷
We are all waiting for Krea2.
I've been coming back here everyday for a week now looking for the KREA2 version. I hope it arrives some day!
Hi, as the best pusyy fixing expert on Civitai, could you do one for Krea 2? That would be awesome.
I second that!!!!!
Need Krea2 🙏🙏🙏
I second that!
Please make a Krea2, thanks!!!!!!!!
Best pussy lora need a krea2 version plz
Krea2 ? <3