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
FAQ
Comments (30)
Unable to generate correct male genitals
Not unable, but not good at it.
I have found the following the best sampler/scheduler combo for decent male genitals:
Sampler: euler_ancestral_cfg_pp
Scheduler: ays_30 or ays_30+
Always use either euler_ancestral_cfg_pp or euler_ancestral as Scheduler.
All these files AIO, they doesnt work in comfyui ?
Of course they work in comfyui.
@6tZ i know it was a stupid question, but everytime i load those files my Comfyui just quits.. and a message up in the right corner says "reconnecting".
so not sure what i do wrong.
@Spectral72 these work amazing in comfyui :)
@Astrolab I guess i am doing something wrong... comfyui just stops and this red message... "reconnecting"
@Spectral72 Is that when you run the workflow? And only if this model is in?
Try one of the other ones then? The non-AIO? Maybe you are running out of memory? What's your VRAM and GPU, and RAM?
@6tZ GPU is a Quadro RTX5000 (16gb) VRAM, RAM 64gb, Intel I7-14700KF
@Spectral72 Honestly, should work, no idea.
Searching seems to show other people are having problems:
https://www.reddit.com/r/comfyui/comments/1kash3e/nvidia_5000_series_video_card_comfyui_still_cant/
Try using the split model version and see if it works.
Good luck!
Any suggestions for settings for your 4.3 AIO model? Sampler, Scheduler, steps and settings for rescaling?
Yes, look at the sample images and their included workflows :)
As usual for ZIT.
I'm using Euler, Beta, 9 steps. / Res Multistep, Bong Tangent, 10 steps.
@KolAR Hmmm. I'm using the same settings I normally do and I'm getting really bad images.
@KolAR Not sure who can answer this because I use Stability Matrix (comfy). Where do I find new samplers? I keep seeing people mention Bong Tangent etc but it's not in my list. Not even sure where to put them if I find them.
@InsidiousOne Just use something else? It doesn't matter too much :)
@6tZ I thought since some samplers work so poorly that maybe the ones I'm missing might work better.
Try them I guess.
I find good results in many samplers, it's not harder than testing them and settings yourself.
I use comfyUI, 12 steps, cfg 1.0 to 1.2, DPM++ 2A, FlowMatchEulerDiscreteScheduler. works great.
This might be a stupid question, but by AIO do you mean that everything is baked in and can run like an SDXL checkpoint on Forge vs running on Forge Neo?
Correct! AIO means it includes the VAE and CLIP model.
So if those tools are configured to run Z-Image Turbo models like this, then it's possible.
If the tools are not supporting this, it won't be.
For ComfyUI, you can run it with the Load Checkpoint node.
@6tZ Damn. Nice. You did what I always wanted to do. Did you do it with python programming or with ComfyUI? That is something that I need to learn to do so I'm curious where to start heading towards. Both I am decently familiar with.
@freestuffpl0x42069 It's all done in ComfyUI.
You can use a "Save Model" node to save model data, or "Save Checkpoint", to combine different sources into an "AIO"-version.
zitc43BF16.bBLq.safetensors is wonderful and only pulls 12 gig ram in Neo forge, this is a winner.
Anatomy seems good for the most part with my testing, but I can’t for the life of me get it to do proper penetration, no matter how I word the prompt.
It's hard to get a good one still unfortunately. Next version has improvements in this area specifically!
@6tZ its quite easy actually. for the most part its easiest with POV shots. Side views are more difficult to do because the model keeps wanting to show the female parts. In SDXL, sometimes it will explicitly perform the act which will produce good angles such as side views, upside-down, front view, etc.
@CH11442 please share the prompt so i can analyze it.
@Melodic_Possible_582589 That is true, POV works a lot better and more consistent.