ALL PREVIEW IMAGES HAVE WORKFLOW EMBEDDED
Some people loved V6 as it was, others didn’t — but don’t worry, you can decide for yourself which one you like better.
I’ve just finished V7! It was retrained from scratch and focuses on reducing skin noise for a much cleaner result. Even though it uses the same dataset as V6, the outputs are noticeably different.
V7 is here! I’ve also added side-by-side comparisons in the description so you can easily compare and choose what works best for you.
Unlocks XXX on Z Image Turbo
Description
First Version
FAQ
Comments (45)
PERFECT, i knew this one was coming, thanks a lot bro!
Thanks for the kind words
Ive trained a few NSFW loras myself and tested most of the ones uploaded here. I dont think we are going to get a decent NSFW image until the base model is released and trained on; which should be relatively cheap, as cheap as SDXL checkpoint training. I do wonder why its taking so long to release the base model. I wonder if they plan on releasing it on Christmas day, the Chinese AI bros do love their memes and troll western developers all the time on reddit and huggingface. But even then, that will likely take a few months. The SDXL NSFW Checkpoints didn't fully become polished until about 1 year and a half later. So its still going to be a while until SDXL is replaced as the NSFW Image King.
Look at my examples
@alcaitiff I was saying that in regards to the vagina, its still deformed in some of the pics.
@fox23vang226 I agree. Not great results on the cocks either - still very deformed
@JazConPlay178 I'm already working on v2 to improve details. But v1 is already a lot better than nothing.
@alcaitiff Its not your fault, its a great lora, the best nsfw lora, its just that we are working with a distilled model is all.
They may be removing celebrities' facial data because the accelerated version can generate a lot of illegal and unethical content.
@wyxzddsjj919 likely why its taking them so long to release the base model. I wonder if they will poison pill the under-trained nudes too like QWEN.
How can I combine this with other character lora's. The weights are combined and the character loses its identity. Maybe Block weights?
Fantastic lora, besides one major issue. It doesn't work well with other loras at all. I tried lowering it down to below .5 and it still causes major distortion for any lora I tried (style, nsfw). Gotta say though, this is really awesome by itself. Thanks!
not an issue with the lora, its issue with z image being only a distilled model. It makes training loras for them difficult and mainly can only use 1/2 loras max. double check your scheduler as well for this one seems to use dmpp_2m_sde/beta
gracias!! lo amo!
I WAS WAITING FOR THIS!
Thanks for sharing. Does this LoRA algorithm have any requirements for the scheduler and sampler? What is the activation cue word?
You don't need a trigger word for this.
For Scheduler and sampler I got good results with:
Euler/Simple
dpmpp_2m_sde/beta
Euler_ancestral/FlowMatchEulerDiscrete
dpmpp_2s_ancestral/FlowMatchEulerDiscrete
To use FlowMatchEulerDiscrete you need a custom node installed https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler
@alcaitiff THX❤️
yo finally a lora that works
Great lora, only major issue I have with it is it can't do innie pussies. Any chance you'll add that in a new version?
oh noes, you'll just have to settle on the vaginas 99.99% of women actually have! 😱
@Narz 👈 That only happens for those who can only get 4/10 women 🤷♂️🤣
@GlowingGuardianGirl with your grade-school joke you just broadcasted you have no idea what you're talking about and this subject matter is beyond your understanding and experience.
@Narz You're very wrong, it just means you're getting the ones I'd never touch 🤷♂️🤡
Most of a time it creates bodyhorror, even when used wf's from example images, but with different seed. Cherry picking lora.
The seed changes very little with ZIT. The prompt is the most crucial part. This lora needs a huge way to get perfect but the first version is a huge leap in the right direction.
This is a proper prompt for ZIT:
## Subject
A young woman with short, dark hair, sitting upright and facing forward.
## Composition & Setting
- The subject is seated, filling the frame from the upper chest to mid-thighs.
- She is positioned centrally against a plain, bright white background.
- Her posture is slightly forward, with her arms resting at her sides.
- The perspective is a medium close-up, focusing on her upper body and hips.
## Lighting & Atmosphere
- Bright, direct sunlight illuminates the scene from the upper left.
- The light creates soft, defined shadows that accentuate the curves of her body.
- The overall atmosphere is clean and clear, emphasizing form and texture.
## Colors & Textures
- The woman's skin is smooth and tanned, with a glistening, slightly moist appearance.
- She wears a red athletic skirt with white and blue trim on the waistband and hem.
- The fabric of the skirt appears soft and slightly damp, clinging to her hips.
- The white waistband of her undergarment is visible just above the skirt.
## NSFW Elements
- Large, full breasts with nipples that are erect and pink, standing out prominently.
- The lower abdomen and navel are visible, with a subtle dip.
- The hips and thighs are rounded and full, with smooth, tanned skin.
- The skirt is slightly lifted, revealing the upper part of her thighs.
- The overall form is curvaceous and voluptuous, with a lusty, inviting appearance.
## Text Elements
- There is no text present in the image.
When I was drawing sexual positions, I also encountered a horrifying effect of twisted human bodies and extra limbs. I reduced the LORA strength to 0.6-0.7 significantly.
@wyxzddsjj919 V2 is in training let's hope it improve the male anatomy
@alcaitiff How much longer? I'm really looking forward to it.
i use grok for make good nsfw prompts, it's a game changer and you can use an image imput: (prompt describing all aspects necessary for accurately replicating the original image.Include specifics about objects, clothing, hairstyles, intricate details, accessories, photographic equipment, environment, lighting, style, body poses, and any other relevant elements, ensuring that every detail of the original image can be precisely recreated:)
@alcaitiff beautiful prompt. i don't know how z-image couldn't interpret that pristinely
@SeoulSeeker Because this version uses an accelerated model, the model itself has issues. Using the UIPC sampler, you can even draw an image in just four generation steps, while restoring the normal CFG values using ZIMAGE (available for download online) requires 40-50 generation steps, making the time to draw an image several times longer than the current accelerated model.
@wyxzddsjj919 V2 is out! It's not perfect, but is better and more versatile. I'll improve the dataset for the next version. Let me try some other ideas and we will improve it further.
This is a great step! however it has some issues as well :D, hoping they will go away in future; the built in celebrety catalugue becomes a plastic skin with this, it also seems it reduces prompt adherance toward lighting
Yes, the great author is training version 2.0! 😍
V2 is out! It's not perfect, but is better and more versatile.
@alcaitiff thank you again sir! will run it and repot feedback. Great work!!!!!!!!!
Hello Alcaitiff, I hope you’re doing well, and congratulations on your new LoRA, which is as excellent as the previous ones.
We are lucky to have someone as talented as you in the community, thank you!
Thank you for your kind words
send some buzz! you are saving z-image :)
Thanks, don't need to send buzz. Post your pictures here I want to see and improve this lora.
Great work as always👍🏻
Thank you for your kind words.