Z-Image is here, and it's pretty damn great. Let's make it more NSFW!
Recommended Settings
Steps: 12
CFG: 1.0
Sampler: dpmpp_sde
Scheduler: simple
FP8 Download
For some reason, when you upload multiple formats (BF16 + FP8) on one model, which they now support, they hide the FP8 under the BF16 model group in the side menu...
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.
Description
FAQ
Comments (30)
7.1 bf16 is amazing! Goodbye 5.0 :D
Wohoo! The streak/curse is broken? XD
7.1 is really really good! I don't mind saying that I have benchmarked a lot of models for NSFW use and previous versions of your models faired OK but not great. But 7.1 seriously lands squarely in the top 5 of the 60+ models I have evaluated, which puts it in the realm of moodyPornMix and cyberrealisticZImage, both of which have been excellent for NSFW generation. Whatever you adjusted and tweaked to get to 7.1, it is working!
Thanks! The key additions here is additional training on penis, and a small dose of Mystic XXX v6. That one added a lot of stability as well. On top of general training on beautiful images from various models.
@6tZ awesome! Keep up the good work
how do you train with it in onetrainer
7.2 is mind-blowing, lost for words just amazing!
yeah right
As soon as I try posing people, this model has a tendency to make them fat and adds leg hair pretty consistently. Like, a LOT. It's fucking gross and I have to wonder what the hell you used to train this checkpoint...
What pose-prompts are you using? Please share a few examples.
LMFAO turbo will constantly do this, focus on weird words, you've probably got one word in there that's messing with you
@teamCentury right. it could be a translation issue. if you do it in Chinese it is a night and day difference. It's like how if you type "anal sex" it always does vaginal sex. Usually "anal penetration" does it or something clinical. Since Ali Baba isn't familiar with western slang it resorts to clinical. (e.g., bald makes a bald person vs no pubic hair). It also depends on how the user captioned when training LoRAs. You can tell when someone uses an abliterated Qwen 3 model because the prompt adherence is good with NLP vs the old captioners that are borderline boolean like Florence2.
Used it in ComfyUI with Clip = qwen_3_4b, and type = Lumina2. Steps 50, CFG 1.3, sampler = seed_3, scheduler = beta. Denoise .85 with 4 LoRA's daisy chained together and the 4xFaceUpDat upscaler for finish. Yes it takes some time to cook, but the results are very crisp and real. Patience is a virtue. Outstanding work really. #contribute
any suggestion on how to use Character LoRA with this? doesnt seem to work super well :(
your character lora needs to be trained on the same base model family - e.g. it should have been trained on some variant of zimage base, zimage turbo, or some distilled model that was based on zimage. if you have a lora that was trained on a flux model, or stable diffusion for example and you try to plug in that lora into the zimage workflow, the results are going to be bad. some models can be very different even if they come from the same base model family, so you might need to re-train your lora on this specific version of zimage turbo, or pay someone to do it for you.
You need to use an original unmodified ZIT checkpoint. NSFW models are highly altered and dont play well with most Loras.
Please ADD Q8 GGUF or FP8 versions of all models!
For those of us that need smaller models to run!
I made an open source tool that is easy to use https://github.com/qskousen/ggufy that helps you make your own GGUF's. It's pretty fast and has very low memory usage. It has a drag and drop UI as well.
@ferretduck Thanks I will try it out see if it work better than other converters I tried and failed :(
I won't have time to do it for all versions.
I do it from time to time on some versions of models. Too much work to do for all.
Ask CivitAI to support auto-convertion of models to other formats.
When making character lora with ai ostris toolkit which model should i choose?
It depends on what model you want the lora to use :)
Just as 6tZ, choose the one you like but its worth mentioning that AI toolkit works best if you train it on diffusers. I wouldn't recommend safetensors, not even sure you can train it on safetensors file.
Character likeness LoRA trained via AI-Toolkit with adapter works perfectly on regular Z-Image Turbo, but fails completely on the base model and breaks generations/performs poorly on the NSFW variant. Any ideas on how to train a LoRA specifically for the NSFW version, and can I use a safetensors model directly from the ComfyUI models folder?
@zairmonibutcew Do, as @JFTC mentioned, you need to do the diffusers model I think, and you need the other files too, in order to train on it. But you can get the original ones, and just swap out the model, you need to edit the config to link to your model, it's a big mess.
@6tZ I've already dealt with it, converted it to the Diffusers format for AI-Toolkit, training has started and so far everything looks promising.
@zairmonibutcew any updates? How did it work out for you? I tried to train on BigLove's merged checkpoint diffusers and it ended up as a garbled mess. would love to hear if it was a success for you.
@Tezozomoctli I successfully trained the character LoRA on the Z-Image Turbo NSFW_6.0_BF16 model without any issues, and the resemblance is satisfying. However, you can see those plastic textures and colors, like something straight off the cover of a glossy men's magazine.