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
Trained for 246.000 steps on a huge dataset. Special thanks to Alcaitiff for their awesome work!
This is an AIO Checkpoint version. CLIP and VAE are included.
FAQ
Comments (39)
Having trouble getting LoRAs to work with this model?
Before you give up, try ModelMergeSimple.
It’s a ComfyUI node, and it often fixes LoRA compatibility issues instantly.
How to use it:
Connect two Checkpoint Loaders, two UNet Loaders, or one Checkpoint + one UNet
Feed them into ModelMergeSimple
Then connect the output to your LoRA loaders
If that still doesn’t solve it, try this proven combo:
Model Loader #1: a model you know already works well with LoRAs > Connect with lora model loaders.
Model Loader #2: this NSFW model
Merge both using ModelMergeSimple
After that, your LoRAs should behave normally,
instead of looking like they were generated by a sleep-deprived potato 🥔😄.
If this works, smash that like button out of pure gratitude 😄
If it doesn’t work… like it anyway... for emotional support, moral victory, and because you already read this far 😌👍(lol).
As long as you use a proper LoRA loader with support for ZiT it should work.
The official one supports it properly now I think.
It doesnt work and shows an error message: AttributeError: 'Linear' object has no attribute 'weight_scale'
@fsh101686 Can you post a link of the lora that does not work with this checkpoint. i will test it out myself and see if i can get it working
19Gb, is it an AIO version ?
Yes, it includes model, VAE, clip, and it's also the full bf16 size, not fp8.
It has much better quality with bf16.
@6tZ, why is v2.0 BF16 labeled as an Illustrious checkpoint? Isn't ZiT it's own thing outside of SDXL-compatible models like Pony and Illu?
The 10fp quality and color comes out great without upscaling. It's respectfully better than the kaiji version. Can't imagine the quality for the full version. I'm still on the fence about this checkpoint as the NSFW is not fully capable yet. Hope more training data on side views and other poses.
Agreed, still improving. It's far from perfect. A lot of body horror still. But it does produce capable outputs, and it was so much faster than Flux got it.
The BF16 is a milestone! It surpassed Chroma1HD in quality. I use it with Huihui-Qwen-3-4b-abliterated and it just does what I tell it to do! It was worth the 7k! Thanks for the great work!
Thank you kindly! Expect more improvements in stability in the future!
Whats a good workflow for this one?
@rpapist1971 Any chance you could share your workflow? TA
hi, you have any camera angle prompts to share? the classic stable diffusion ones are having a hard time working out and i also use the camera prompts generator that show a lot of the detailed prompts but its really difficult to get them to adhere. do you think this was a training issue where the checkpoint didn't have other angles?
https://www.reddit.com/r/StableDiffusion/comments/1pb02yu/camera_angles_comparison_zimage_turbo_vs_flux1/
I honestly don't know about camera angles or the training of it at all. Haven't had time to consider it.
For more camera angles, you can always ask an LLM to make a wildcard list for you.
Can we get not an AIO version, as disk space is quite important.
-
Sure, I can upload a separate one.
@6tZ bf16? Could you make fp8 scaled/rounded?
@qek Yeah I can convert to all, just need a bunch of time to upload. I'll use the same sample images, too much work to create all new for each version. But the models are working fine, slightly less good but good.
I'll skip FP16 I think and just do BF16 and FP8
BF16 and FP8 is up for v2. Sample images are from the regular model as mentioned.
@6tZ haha, qek, they're paywalled. I had to download the checkpoint and extract its transformer myself
@qek Yeah, of course anyone is free to do that :D
this version dont need vae and load clip, yes?
yes
Correct, the version 2.0 includes it as a checkpoint.
Results got better with default vae and clip loaded on my machine somehow, especially with genitals.
@AnonymousAAA Oh, interesting. Gonna have to do some comparison testing with that. Thanks for the notice!
looks amazing, hopefully waiting for gguf versions
I just uploaded fp8, would that work for you?
@6tZ check again. it complains that clip and text encoder are missing, for a checkpoint
@okayd Which version are you using? The AIO or diffusion? Only AIO has VAE and CLIP.
You can convert using this guide..
https://drive.google.com/file/d/1QdmtXijMROZC8nTpKPJGQ_Xlt8OIJK6U/view?usp=sharing
If there's an error, follow this information - https://github.com/city96/ComfyUI-GGUF/issues/381#issuecomment-3666642534
Okay, now we're getting somewhere. Surprisingly good considering this isn't the base model.