Z-Image is here, and it's pretty damn great. Let's make it more NSFW!
Still a work in progress. It can output good genitals, but not always. About 3/5 images are good, working on penis stability specifically for future versions.
ComfyUI: Models should be placed in diffusion_models, and uses the official workflow with the Load Diffusion Model loader, and loading the VAE/CLIP models separately. Models named "_AIO" should go in the checkpoints folder, and uses the normal checkpoint loaders.
Recommended Settings
Steps: 12
CFG: 1.0
Sampler: dpmpp_sde
Scheduler: simple
Description
Version 6.1 has a few issues with anatomy. It can still do great outputs, but it's a quite a bit less stable.
FAQ
Comments (39)
I can't get this to work with character loras at all. Am I doing something wrong or is that just a limitation of the model?
Can't do it for now as far as i know. Maybe training on turbo? We need to have OP confirm
@mobdik17378 @mastodonlopez351 Z-Image Turbo is weird, with some models it works well to combine with other LoRAs, other times it just breaks out completely. I think it's a problem with the architecture, but I'm not an expert. Maybe it's already solved somewhere? I don't have capacity to read all the research or follow all the news I'm afraid.
If someone has any advice or tip on how to make the model work with other LoRAs, I'm eager to listen!
I get it to work with loras trained on ZIT but it doesn't respond well. This model breaks up as you get closer to 1.0 on the strength for a lora.
@6tZ the pmaster team figured it out, their ZIT model works with character loras (as long as its trained on ZIT)
@EchoSingularity you can share the link to pmaster, thank you
@6tZ Got it working with regular ZIB models. You need to train on ZIB and increase the strength for character loras, this works with ZIT and ZIB afaik
is the 6.1 version as variable as the original BF16 version ? is it possible to purchase the lora of version 6.1 ?
I don't have a LoRA version of this. It's an amalgation of lots of trainings and block manipulations merged into one checkpoint. You could perhaps extract it into a LoRA, but it would not be exactly the same thing.
Damn you are so fast! Cannot wait to try the 6.1 as your ZIT is currently my absolute favorite (and I tried ALL checkpoints avaiable on civit).
All checkpoint!? Damn, that's hard work! I salute you! Thanks for the kind words, I hope you like 6.1 as well.
It can output some real stunners, but is a bit unstable.
@6tZ Yeah, even thought about starting some articles for ComfyUI and checkpoints. But I am helly busy writing my book rn so that will have to wait.
(I am also using your checkpoint a lot for inspiration to my book. If I have a place or character in mind I generate an image and use it as a base for a better description and it really works wonders)
Could you please advice which of: load CLIP; VAE and lora you are using? I tried my old preset for this new model but it's generating only mess...
@um944 I am using the wf, vae and clip provided by SubtleShader and play around with Aura Modeling Shift, however I remove the link to the upscaler and only let it impact the latent image.
Thes rest is prompting and finding the right seed. I have seen people use the "old" style prompting a lot for ZIT ("1woman, blonde, sexy...") and while it works for some models, even then they often respond much better to lengthy descriptions and short phrases in natural English or Chinese language.
Hope that helps!
@um944 oh and I rarely ever use any loras. Only for Klein Anime to photo workflows!
what do you have as training sets? what n*s*f*w stuff does the model know? also do you have specific training sets for asian caucasian etc? i did test and atleast african-american asian japanese korean slavic and eurasian work very nicely.
these i have already found working:
d*ck, p*ssy, br**asts, mis***sionary, dog***gystyle, blow*j*b, cow**girl, an*l, han*dj*obs, ti*tj*obs, twe**rking(as**s shaking), dancing, bigb**lackc*cks, grou**ps*x, any size bre***asts if you prompt it correctly, pov view of almost any s*x pose.
what i found broken on v6.0:
nipples get wierd very often, like huge looking nipples like a p*nis, extra geni***tals, d***ck is the wrong way around sometimes, sometimes females look like transgender if you dont specify age and bodycomposition properly.
but overall this is the best n*s*fw zit i have tested.
also those who have problems with saturation: there is tons of color adjustment nodes on comfyui... upscaling is also a thing so quality is no issue.
(damn cant even type here anymore without getting censored :D )
I have all sorts of training sets for this and NSFW.
I have a general 24k decent quality training set that I use to get foundational NSFW in. It's not the best quality, but it helps to get good nudity.
I have male-specific training sets, but they are less detailed and has much less content. I need some better datasets, so if anyone has any to contribute, hit me up.
There are poses and such included, but I don't have huge datasets of them.
I have trained for a few hundred thousand steps to get foundations in, and I need to work a lot more on multiple people. This is where it's weakest.
A lot more training is needed to avoid body horrors. I still get them all the time. With version 6.1 it's worse than before, but I still really like the good outputs from it. Will focus on stabilizing this for the next version.
The amount of body horror models I have on my drive.... I need to train a checkpoint on just that content.
Great looking stuff, i think it would work very well with my realistic snapshot lora, care to try it out?
Shouldn't you be the one to try that out?
I gave it a spin but the models don't work well together. ZiT does that.
I have been working on an amateur model for a bit, but my results are pretty bad, yours looks much better so far.
@6tZ not wrong but i just asked
@6tZ I got the result magically before they fucked up their z-image-turbo trainer for a month, then i spend 20 dollars on ai toolkit on runpod and havent gotten a better result since, idk i think i was lucky.
Interesting. Would love to learn more about this all.
Your models are excellent, and v6.1 is no exception. It has handled complicated prompts with ease, and the output quality is incredible, thank you for all your hard work!
Is this trained on weiners ?
The more useful question is: What percentage of the training is female and what percentage is young attractive male? That's what people need to know when considering a sex checkpoint. If it's mostly or completely female (or females plus penises) that's fine. Just let people know so they can make an informed decision to take the time (and download cap, for metered connections) as to whether or not to download.
Your question appears to have already been answered because if you look at the samples you'll see there is a penis. Are you asking about women with penises attached?
my own hole is exit only (a joke, haha im so funny), but tested since you asked:
weiners do work on any hole or no hole. sometimes weiner is just upside down, but then generate new image to fix it. weiners on female holes and weiners on male holes work fine. weiners on raccoon holes too if that is what you like.
also i suggest you use "qwen3.4 abliterated" (check from huggingface there is the model) with this, so then the image generator and text encoder are both uncensored. best results.
It's mostly women. It is trained on cocks too, but it performs much worse due to bias.
I aim to improve, but it's a hard challenge!
@6tZ Understood - looking at a woman is great, but sticking a weiner in her is the ultimate goal. LTX does not really know what a weiner looks like, and as you note, there is a huge training bias toward just female anatomy. Men showing up with manginas, etc. There have been very few models outside of Pony that have incorporated male anatomy into base model training. I can only assume that's a huge endeavor.
To anyone writing comments:
Sorry for not responding, I get 30+ message notifications per day and the notification system for comments is broken, so I cant click to go to your comments, and I have to manually find your message in this sea of comments. It's just too much work. Please create a new comment if you require my attention. Or don't and maybe I'll see it in the future.
This is not the correct prompt input field!
every time I try it is a huge furry ball. Good luck!
Please add FP8 for v6.1 !?!
Sure thing!
can't seem to get this working with character lora and realistic snapshot? :/
For me it's not just this model. Every LORA characters I trained will only work with base Z-turbo
@bendead69 That's a you problem, this isn't helpful