New KREA 2 Version
Faceless dataset (doesn't influence face), can be used with character lora easily.
workflow embedded in all images.
--------------------- V3.0 ---------------------
A lot of effort went into this — hope you guys like it :)
This V3 release includes several major improvements. I carefully reviewed the feedback I received and addressed the issues present in the previous LoRA. Updates include:
– No more color fading
– Faces are no longer altered
– Improved genital anatomy
– Greater overall versatility
– A fully reworked dataset
– A new training method
– Support for both shaved and hairy anatomy
– More body-type diversity (not limited to skinny women)
– More consistent close-up results
Use this lora between : 0.7 - 1
recommended triggers : shaved pussy, anus, hairy pussy
Sampler : Euler / beta - 10steps
worflow embedded in each pictures
--------------------- V2.0 ---------------------
improved high res dataset
no more faces in dataset
adjusted training params
better skin texture and genitals
Recommended settings :
Lora weight 0.7 - 1
Full workflow embedded in the example images.
First try on training a z-image lora.
Keep in mind this is trained on the distilled version and not the base model (not released yet).
Once the base model is released i'll re-train and update the lora.
No specific keywords but i recommend using words like : naked,exposing her pussy.
Description
Improved dataset, better fidelity
FAQ
Comments (111)
Can i upload this somewhere else and credit you?
Sure, can you send me the link when you did it plz ?
Ну пошла жара. Доставайте свои баклажаны. Готовьте вареники
Any idea when the base model will be released? Thank you
According a comment posted by a member of the dev team on github, base and edit models are gonna be open source and a demo will be released very soon (end of week-end possibly).No more informations about the release of the models
@aip0pp Great, this looks promising, thanks
nice, maybe some pubes>? win with hairy 70s style bush.
there is no body variety here...they all look the same weight and height...totally shitty
You can prompt for it btw, look at my pictures, you can make obese women if you like.
Waiting for your lora to compare. Thanks for the constructive comment :)
@fox23vang226 but you didn't use this lora in those obese gens...
I just want curves, not this transvestite futanari new age shit...its making me mad, there is not a single woman with curves, wide hips in these examples...wtf !
@MatorinaDaddy Make your own Lora then? oh wait
@MatorinaDaddy i have a special surprise for you at the end of the examples :) hope you like V2
Zimage knows many specific celebrities, and this Lora has undergone drastic facial modifications.
Thank you I will take a look
Currently, the impact on the face is greater than the impact on the pussy.
Ok I'm taking note for the future version. Thanks
@aip0pp you can solve this by using reg dataset.
the lora will learn that's another kind of faces are normal.
@luisa_pinguin I'm working on a v2 with a better dataset to not alter the faces
Hello... I'm getting an error on my end... Lora isn't loading. In your workflow, Lora indicates version 2. but we have 1.1. (sorry my english is bad)
Thank you for your work!
Select the lora you downloaded in the list in the node
Qwen-Image LoRA Loader showing not 1K warnings only One ;)
@aip0pp Yes, of course, I did that. But I get this error page, like when you install a LoRa that's not compatible with the model.: lora key not loaded: diffusion_model.layers.0.attention.to_k.lora_A.weight...
@charlie_perso are you using fp8 ?
@aip0pp no.. bf16
@charlie_perso I know there is a special node to load the lora in your case to avoid errors but can't find it anymore, I've seen it on reddit
@aip0pp Yep ! solved! in Comfyui, look at CRT nodes! thanks!
I like it, it works well on SFW / clothed pictures as well and gives them a great touch
Face bias.
Face bias everywhere.
.. And body type/age bias 🗿
Why a nudity lora is alway damn hard to make?
Can we all share a main dataset that is proven to work and join our forces? FFS same story for every new models.. is so depressive to watch same mistake over and over again
Can't wait to see your version of the lora :)
@aip0pp So, I like the lora a lot, and I don't think what he said is untrue or your fault. I am quite sure we can all agree that it's very hard to get specific body types. I see the same issue in your own examples. But I guess this can be fixed once more body type lora releases by mixing them both.
@p_p because majority of the people trying to make these Loras for new models aren't trying to make "The best" they are trying to make the first.
They all have a fundamental misunderstanding on how to make Nudity/NSFW Loras or concept Loras as a whole without also training in significant biases. Most of the issues lie within their training datasets, but there are still also issues with their training settings.
Most of the people I know who would have jumped on this and made a fantastic one all stopped trying to keep up with newer models because they no longer had the required hardware, cant afford to rent the hardware, or gave up on Lora training entirely.
I have college finals coming up but I am honestly going to try to throw an attempt or two at a decent NSFW Lora for Z Image.
@willywizard That should be great. Version 2.0 of this one killed a lot of my prompts. It alters the body but also ignores my hairstyle, colour, body style, etc that works very well with the base model.
@MrCylion A lot of these people opt for generalized multi concept datasets without any cropping, meanwhile for a proper NSFW/Nudity lora, the best bet is to use a smaller dataset of only around 100 images and cropping them to the content of the concepts in which you plan on train into the lora. The fact that their dataset contained majority full body images likely with a low amount of variability is what caused the significant biases and is interfering with your prompt.
@willywizard Thanks for the infos, i'll try to improve my dataset to achieve better results and i'm waiting for the base model and not a distilled one. v 2 is using cropped image dataset but maybe not enough close up or bad description, i'm still learning and trying to improve :)
significantly alters the character, not well trained. Would advice against using it with Characters / Loras
for some reason I get errors on a fp8 version of z-image and the lora is not loading, might not be an issue with the lora but could be the fp8 which is using a special loader as well...
I know there is a special node to load the lora in your case to avoid errors but can't find it anymore, I've seen it on reddit
Like "lora key not loaded"? In my case CRT-nodes Lora loader for z-image solved this issue.
@palladium107 For me it was a complete comfyui update that solved it, no more mismatch errors now.
I find the comments a bit harsh. This is the first Lora that's more or less a generalist for female nudity. It might be biased, but it's a good start . With a light weight, like 0.3 or 0.4, it helps me a lot.
There are already at least 4 or 5 other LoRA's for producing women with exaggerated curves and bimbo faces, and nobody complains about it.
Of course, we're waiting for other, more diverse LoRA's. The model needs to learn more, and that won't happen in two days, nor with just one Lora/Finetuning. Especially without the complete model for now.
Anyway, thank you very much, it's a good start and it's promising. Looking forward to the rest :)
@Error666
Thank you very much :)
as you said it's a first iteration. I'm already working on a better lora with a high res dataset and training in high res.
But don't worry I'm not taking into account comments when they are not constructive and/or coming from angry people :) and the base model isn't even released yet.
Some people seems to forget that all of the free lora comes from passionate people that take time and money to train for the community and for free. Thanks for the support !
Give the new technology some time—miracles aren't achieved so easily. All we can do is support the author in continuously optimizing it. Thank you once again for your work.
These results seem to be coming together much faster than on Qwen. I'm guessing the number of parameters helps in that way. Plus the loras are smaller.
yeah, iterations are way faster, like 1s/it on a 4090
It's a good effort, but in its current state 2.0 the images look worse.
Have you tried DPM++sde/simple ?
@aip0pp DPM++sde / Simple does not work with Z-model.
@HonZuna copy the nodes in one of my examples, it works with more steps
Yes, I tried multiple sampler combos. The lady parts are more defined, but it changes the lighting to be really flat and uniform, which makes it look like low-contrast SDXL. Do some side-by-side comparisons in different lighting conditions and you'll see it.
@mweldonsd594 will check and take it into consideration for the next version :)
Can you please share your training params?
@desm0nt i followed this video : https://www.youtube.com/watch?v=Kmve1_jiDpQ but using 1280 resolution.Depending on your lora (details etc) bump the steps (don't hesitate to use 10k steps, but check along the way for overtraining artefacts)
and thanks for the tip :)
is the dataset made with synthetic data or real photos?
It adjusts characters, and the generated image often has the head out of frame, requiring prompt adjustment.
yes i used a dataset without faces, you should use word like smiling or detail the face to have faces in your images :)
Props on being pioneer on this front! Wish you the best head
this lowers the quality of the image by quite a bit making it not worth using Z at all
Image Quality seems pretty solid to me, you using any other Loras on top?
Same here, it is also very washed out. I can't seem to force depth of colour. The washed out look scales with the strength
@Questor i'm working on a new version to avoid this
@aip0pp I have to apologize. Its actually fine, but I was using a random shuffle workflow that was the real cause of the problem. Keep up the sumptuous vulva work.
@Questor no problem, glad it is working :) thank you
“Steps : 25”绝对不对，伪影非常严重，这个模型是steps越少伪影越少、细节越少、越“光滑”，steps8、11、13效果相对比较好
Good thinking with 'No faces in data set' :) any support for darker skin types aswell? maybe that would need a seperate Lora?
I'm waiting for the base model, and yes i have to include darker skin as well :)
I have also been trying to train a similar more realistic body. Your's came out a bit better than my first attempt, which I'll probably publish in a few days.
How many images did you use and how many steps did it take?
I am using around 350 images at various distances from zoomed in to full body. It took me until about 5500 steps until I got fairly more realistic looking bodies, and it took until 9250 to consistently get good pictures.
Also what purpose was removing the faces? The faces in v 1.1 look a more realistic than 2.0.
Hi, around 40 images, 10k steps to have details. I removed the faces because it had too much impact on the model, generating the face it was trained on and not being enough flexible.
@aip0pp Thanks, so it looks that around 10k steps might be normal for it being distilled. I thought it was a bit too high on my test.
Why are the genitals also so average or deformed? Any idea how to get a better vagina or insertion combos?
Off to a great start!
Works ok and makes the nipples and genitals less bad, almost tolerable.
If you want to make heifers or (I presume some other odd body types) then it waters down those instructions.
looking forward it~
Hello. Could you please tell me how to install the LoraLoaderZimage node from your workflow?
@alezzby Hey, sure here is the github repo : https://github.com/PGCRT/CRT-Nodes
Thank you, when I load ComfyUI 0.3.76, it says that everything is fine. D:\SD\ComfyUI_windows_portable\ComfyUI\custom_nodes\CRT-Nodes,
[CRT-Nodes init]
Registering custom model paths...
... 'ultralytics_bbox' path registered.
... 'ultralytics_segm' path registered.
[CRT-Nodes init] INITIALIZATION COMPLETE.
However, in the workflow itself, node LoraLoaderZimage is red :(
@alezzby did you refresh the page ?
@aip0pp Thank you for your response. Yes, I have updated it. I have also tried installing this extension using the comfyui manager and the git terminal.
Your work is good, are you going to try to train on the de-distilled version of ZIT released by Ostris?
Thank you :), the v3 is trained on this de-distilled version already ^^
Hi. I just bought and tried V3 and it's having a huge impact on the face, quality, etc. with identical generation parameters. Even at lower strengths (~0.7). The pussy is far better with your lora, but everything else about the image is much lower quality and less attractive than without it. Any thoughts?
Hi. Strange, did you try my workflow ? Embedded in the images. There is no more faces in this version of the lora
Seems to work very good. But only produces pale skin for me. Please train with more variety of skin tones :)
@ardanlilith622 Thanks! You’re right—I’m planning to rework the entire dataset to include a wider range of ethnicities, more diversity in pubic hair styles, and better representations of breasts in different sizes.
@aip0pp could you please train on a diversity type of areolas, they look all flats, ik lot of womens have it like this but its better to have pumped areolas not too much just for aesthetism.
"Crowd is very diverse (arabic and black mostly ,age, body, genre)"
wow that was interesting, i actually wonder why you typed that 😂😂
For comparison, I created a workflow where everything was the same, with only the difference of simultaneously starting with the same model and going through this Lora in the middle stage, and the result was that the overall color/body/face details of the characters were completely different. It almost reminds me of qwen.
Maybe it do somethng but i didnt noticed
Good work !! And nice workflow but i can't use the upscaler idk why
male nudity when?
This is the best ZIT LoRA I've ever used—it doesn't alter the face, and the body structure stays intact too.
Great work! The only Lora in all of Civit who doesn't affect the face.
can someone tell me how i can use this lora
hey, workflows are included in the images, just copy paste the nodes
@aip0pp how can I combine this lora with other character Z-Image lora? any workflows? should I connect sequentially? what should be the weights? also will this lora lessen the likeness of character lora?
@aip0pp i copied the images and put them in comfy ui as the nodes but nothing happened, sorry if im bothering im just new to all this
@djonjajaja201 no problem we all started at some point. you can click on an image and click on download top right.then move it to comfyui to load the workflow or click on the i at the bottom of any image and click on copy on nodes and then paste into comfyui :)
@aip0pp Thank you very much i watched a lot of tutorials the last couple of days so im way more informed now!
Nice pussy, brother!
How it's "better"? For years models generate genitals worse than hands, lol
Nipples often come out pretty bad quality, flat and malformed.
Why nipple always missing, only the areola visible, step 10, euler, beta