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 (19)
like this so much, it'll be great if there are tips about the improvement of different version
I would love your help. Please write an essay about the differences you feel in the different versions.
@6tZ best creator replies on the site 10/10 would read again (and did)
Even with 5gb checkpoint size, it takes up to 26minutes in 6gb cards (batch=2, 10 steps, res_2s bongtangent), 1024x1024. I use the clip Qwen3-4b-Q6_K.gguf, 3gb size. In RTX3060, it takes just 2min. Is there a way to accelerate this even more for 'potato' cards? Like using a 'turbo paseer' or 'dmd' Lora for Z-image? Thanks in advance!
It's Turbo by default, hence the name "Turbo". This means it's meant to generate in few-steps. It's fine-tuned to give good images quickly.
But, you're already doing 10 steps, so you can't go too much lower than that I'm afraid.
@6tZ Thank you for the proper response! Well, provided many still have 'potato cards' before Ampere architecture (or AMD/Intel cards), we can either try to use an even more optimized Qwen3-4b, say, Q4, or use a fewer res, like 800x800 or even 640x640, then upscale it using a SDXL/Pony/ILL checkpoint that has 'akin' styles. (so I can render many at 640px then choose the ones I want to upscale).
@schsch Yeah, I'm astonished potato-PC-folks are even able to use these models in the first place! The community is truly astonishing in bringing these ones down to being runnable.
But 30 minutes for an image is a bit, yikes.
I guess you really need to half your resolution, and run like, 4 steps, and see if the thumbnail (that's what it would have to be), looks good, and then store that for full-resolution queue overnight?
I would do something like that.
Or save save save for a non-potato.
Hello, i;m using the workflow inside your posted images. Is there a way to get live preview of the images you are making.
Well, I tested the 9.2 model, I'll say this, the model is good, there is no strong oversaturation, details, promta listens better
Hi, big fan of your models. I have used your 9.0 and 9.2 ZIT FP8 models, and they are the best. But.. instruction following can be improved.
Any breast sucking scene, especially when combined with sex, is a hit or a miss. Sometimes it comes out great, most times it does not. Please check deep breast sucking LORA for SDXL/Pony. Unfortunately there isn’t an equivalent for ZIT. But I hope we can improve the support in ZIT itself.
Also, missionary sex position could be improved. Other sex positions don’t work great. Sex against the wall especially. Lotus sex position too. Yeah one more thing, sex position where the girl wraps her legs around the guy also needs more improvement. It doesn’t always wants to follow that.
Thanks for testing these positions.
I'll add it to my todo-list.
@6tZ Much appreciated.
@6tZ I tested your 9.3 model and it’s positive strong point is much natural looking excellent images. But in my humble opinion, instruction following was a bit better in 9.2. Especially in the cases I mentioned before. Keep up the exceptional models coming. They are the best.
anyone else getting a different hash than the one posted?
Posted where? The Civit AIR thingie?
the hash above Sha256 posted as part of the download to verify. First time I've seen one different to what stated.
@creatorsblock huh, maybe civit is doing something weird. Which model version?
I tested v 9.3 and I'll say this, it surpassed 9.2, the photo is clear, the details are high, I really liked it, but also according to the father-in-law, thank you for the beautiful model, it seems 9.3 has become my favorite model
This v9.3 model is really successful, realistic, thanks a lot!