First Release – Version 0.1
(EDIT: Since I have received quite a lot of feedback, I just wanted to write something about this. The most important thing first: there will be no sex poses or impressive genitalia. I tried a lot, but I could not get anything more than mutant genitalia, and I am also a bit tired of generating sex (yes, he really said that).
I also found out that Z-ImageTurbo handles nudity quite well. Therefore, this will be more of a semi-realism Model with a focus on adult aesthetic nudity.
This merge was made only with women in mind. There will be no male version, because I simply do not enjoy generating men. Sorry about that.
Version 0.2 is already finished, and has fewer problems with fingers, but the style has also become more refined. I'm at the showcase right now.
Because some people follow me specifically for my non-NSFW content, I would like to apologize here.
This is an attempt at an NSFW checkpoint for Z-ImageTurbo. I mainly just wanted to test how well it works. The nudity here is focused on breasts, it shows little to no genitals, the nudity that is present is fairly simple and works best in portrait, and it has not learned any sex poses yet. It only knows large breasts and mature bodies for NSFW.
(btw, it works great even without nudity if you like the style)
• bf16/fp8/fp16 Diffusion Model
• No CLIP and no VAE included (Ask me if you need help with that.)
• Recommended settings: CFG 1, 8 steps (max.15)
• Sampler: Euler A, Scheduler: Simple or Beta (Beta highly recommended)
• Sample images are not upscaled and no Hi-Res Fix was usedOriginal ComfyUI Models: Link (here you can find CLIP and VAE)
Description
FAQ
Comments (15)
我正在为没有合适的动漫模型发愁，不过好在找到一个令我满意的作品
Thank you very much for the feedback!
I'm glad you like it 🔥🐎
If you want it to be able to do NSFW stuff, you need to swap "text encoder", in this case actually LLM, for one that can do it. Otherwise training wont help, it will get blocked at encoder level.
What text encoder should I be running in my Comfy workflows?
@97Buckeye I don't know what's going on here tbh, z-image turbo is already pretty good at nsfw stuff as the images I've made show I think, but maybe I'm just too vanilla - my biggest issue is it whiffs on vaginas and penises, and doesn't understand many positions but loras fix that. If you aren't happy with the regular qwen 3-4b text encoder, you could try an abliterated one like huihui-qwen3-4b-abliterated-v2 for less censorship.
Hey! First off, I’m just someone who enjoys generating images, not an expert or anything, and I’m grateful for any tips.
The thing with the LLM is a good tip and I’ll give it a try.
The problem was that the style itself doesn’t have any NSFW, and after the merge the nipples looked like cornflakes or stalagmites. That’s why I thought I’d need an additional nude LoRA based on a few idle poses. I personally don’t really generate NSFW, and because of all the NSFW LoRAs out there I thought it was normal that it couldn’t do nudity.
@VisionaryAI_Studio Default LLM while not heavily censored still might do random refusals, which is why NSFW in general is a bit iffy with Z-image. Apart obvious lack of training on their side.
If for some reason NSFW doesnt work after your LoRA, it might be overfit a bit. Maybe. Not sure there.
@Mescalamba
Regardless of NSFW - is there any guides how to make my own CLIP models from LLMs? I am running some locally and would like to see what differences some of them could do, however I wasn't able to find any results on how one could achieve this on ComfyUI or convert/tune/or even fine tune for example ollama llm models to clip models for image generation usage. I somehow got understanding that this is what you meant. I am quite eager to learn if this is possible - seems like I am spending more time testing and tinkering with workflows than actually generating any images.
@Crampton Thanks for the suggestion! I'll also try https://huggingface.co/fedric95/Qwen3-4B-unc as this feels like an "uncensored" version of the original checkpoint
I like what I'm seeing here. Looking forward to what you can do with different positions and genitalia.
Unfortunately for me it seems to sameface a little too much. Somehow the eyes and face shape always come out similar. Maybe less face focus in the training data and more on body and positions (both genders)?
hehe, the thing is that I love this face ^^
I’ll keep that in mind.
So, first of all, this model has cool style! ( I use fp8 version ) I really like it. So far I have tested most of all z-image models, and this model one of the best, because this model doesn't ruin your output with excessive noise, as if they took a small picture and enlarged it. Even without most of the types of NSFW it is great. Hope to see improved version of this model. Thank you!
Hey, thanks for the feedback. I'll keep working on it, but it's very time-consuming to convert this style into more NSFW, and I'm very busy with my personal life at the moment.
@VisionaryAI_Studio Got it!