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 (22)
What's the difference between 2.1AIO and 2.0AIO? Any improvements?
Not just any improvements, you gotta add the M
MAny improvements!!!
@6tZ what are they tho lol. must be good if they're worth 3k more
@TiffanyHeat @AnonymousAAA It's improving the photographical style of the model, further blending of NSFW LoRAs trained for this (my own and Mystic-XXX v2). Basically improved genitals is the key/goal.
Model preview images are of course cherry-picked, so it was always going to be well-presented. But think of it like NSFW stability. First version had a 50% cut-rate of images being bad, now it's less.
@6tZ I assume 2.2 have MAny improvements as well?
@AnonymousAAA You got it! A lot better skin details and general stability.
Where to get dpmpp_2m_sde_beta ?
It comes with ComfyUI, it may have a slightly different name in other tools.
@6tZ In ComfyUI I cannot choose it.
@Sad_Neighborhood875842 Why? It is dpmpp_2m_sde, the scheduler is beta
@qek NONE of the generated images look similar to the presented ones, even if I add the exact parameter. All are totally different. I have setup dpmpp_2m_sde with scheduler beta.
@Sad_Neighborhood875842 Yes, I had some problem with workflow savings in these images. It will be fixed in future versions (though not all images from them).
The included workflow is ruined and cannot be used.
https://civitai.com/images/116457346
Use this one, it's working fine for me, I can reimport it into comfy and it has the workflow.
@Sad_Neighborhood875842 And yeah, sorry, I said the wrong thing but @qek is correct of course. the _beta part is the scheduler.
It was recommended in an article on civit to use dpmpp_s2_ancestral with the FlowMatchEulerDiscreteScheduler. So you can try that as well. That's a custom node you need to install to get the scheduler.
@6tZ It's civitai's bad, it uses the same field for samplers and schedulers like old versions of SD WEBUI, and it has a poor selection of samplers+schedulers. I am sure the civitai devs will never fix the problems
@qek Yeah they are busy fighting for survival I guess, no time to think about luxury features. There are so many low hanging fruits.
Can this be used with a character lora yet?
As long as you use a correct lora loader that handles z-image appropriately, yes.
Alternatively, you can mutate the layers yourself and re-save the character LoRA.
https://github.com/shootthesound/comfyUI-Realtime-Lora
Inspire Pack's block weight lora loader seem not to work with Z LoRas
@6tZ what do you recommend for training a compatible character Lora? I've been using mostly default settings in ai toolkit with mixed (bad) results. They work well in vanilla z-image, but not this one unfortunately.
@uncropped Hmm, I'll see if I can fix it on the checkpoint side for future versions. I have an idea.
Regarding how to train compatible, I don't know, normal settings in ai-toolkit should work. Didn't try myself, but I've tested on one and it worked with proper z-image LoRA loading (where it mutates the layers).
@6tZ I've found character Lora works pretty well if you use Qwen3 4b Abliterated as a text encoder.
@uncropped Huh?! Interesting, how the hell does this work any differently?