Z-Image Support

[Acly]

Z-Image

Last updated: 1.47.0

Features

• Z-Image Turbo
• Z-Image
• Z-Image Edit (not released yet)
• GGUF
• Inpaint model
• Control Layers
• Managed Install
• Cloud

Requirements

The managed install can run Z-Image models since plugin version 1.43.0. Automatic model downloads have been added in 1.47.0.

For custom ComfyUI installs remember to update:

• ComfyUI
• comfyui-tooling-nodes
• ComfyUI-GGUF (optional)

Model Download

Diffusion Models

Download one of these. They go into the /models/diffusion_models folder.

• z_image_turbo_bf16.safetensors
• Alternatively you can use GGUF for smaller file size: huggingface.co/jayn7/Z-Image-Turbo-GGUF

VAE

Z-Image uses the same VAE as Flux 1, so if you already have it you can skip this download.

• ae.safetensors → /models/vae

Text Encoder

Z-Image uses Qwen 3 as text encoder. Note that it is different from the one used by Qwen-Image. Make sure the filename has "qwen_3" in it (it can be longer).

• qwen_3_4b.safetensors → /models/text_encoders
• Alternatively you can use GGUF for smaller file size: huggingface.co/worstplayer/Z-Image_Qwen_3_4b_text_encoder_GGUF

ControlNet

To enable control layers, download the control models below. Note that they must be placed into the model_patches folder (controlnet folder will not work!)

• Z-Image-Turbo-Fun-Controlnet-Union-2.1-lite-2601-8steps.safetensors → /models/model_patches
• Z-Image-Turbo-Fun-Controlnet-Tile-2.1-lite-2601-8steps.safetensors → /models/model_patches

"Tile" model is for unblur/upscale, the "Union" model for everything else.

Customizing the timestep range where the model applies is not supported yet, so using it comes with some quality degradation. Follow up with a refinement step that doesn't use the control model to mitigate the effect.

Additional information

• There is a "Z-Image Turbo" built-in style with recommended settings
  • Sampler: Euler Simple, Steps: 9, Guidance: 1
• Z-Image understands prompts that use detailed, natural language in Chinese or English.

[ETCore7]

Very nice, thanks.

Below are some quick results with Alternative Euler-A, Steps 10, CFG 1.0, 768x512, model z_image_turbo_bf16.

--
PS, to new users, you can ask AI like CoPilot to write a base prompt with 'wildcards' so the plugin will generate unique prompts each time. It's a nice way to see variations.

[fukc-gihtub]

Prompt following leaves something to be desired.

Alternative Euler-A, Steps 10, CFG 1.0, 1024x1024, model z_image_turbo_bf16, seed 1, prompt "A small angry squirrel pursues a big frightened fleeing cat."

[ETCore7]

Z-image likes detailed prompts on what to make exactly, the encoder has difficulty to interpret vague terms like 'pursues'. Ask AI like CoPilot to write a "script" for you.

Example prompt:
Scene: A dramatic autumn garden path, scattered with orange and yellow leaves.
Action: A small, furious squirrel is mid-air, leaping aggressively with claws extended and tail bristling. Its eyes are locked on a much larger, terrified tabby cat fleeing ahead. The cat glances back in panic, ears flattened, mouth open, legs stretched in full sprint.
Composition: The squirrel is clearly behind the cat, closing the gap. The cat is ahead, off-balance, with motion blur and flying leaves emphasizing speed and fear.
Emotion: The squirrel radiates rage and determination. The cat shows pure panic and desperation.
Style: Cinematic lighting, low-angle perspective, shallow depth of field. Emphasize dynamic motion, emotional contrast, and predator-prey tension.

Result:

--

--

The "wildcard" that CoPilot suggests are options for the generator. "Picture of a cat" generates a cat, "Picture of a {cat|dog|elephant|squirrel}" will randomly pick an animal. CoPilot can generate a base prompt with multiple wildcards. You then can run some batches and see what images fit your need.

Example prompt:
A {squirrel|mouse|duck|frog|hedgehog} {furious|determined|mischievous|gleeful} {chases|lunges at|sprints after|bounds toward} a {cat|dog|goose|rabbit|fox} {terrified|panicked|startled|confused} {flees|scrambles away|bolts|runs in panic} through a {snowy|rainy|foggy|sunny|windy|overcast} {garden path|forest trail|city alley|grassy field|suburban street}.
The chaser is behind, mid-pounce with {claws extended|tail bristling|mid-leap|eyes locked on target|mouth open}, while the runner is ahead, {ears flattened|glancing back|legs stretched|tail raised|fur standing on end}. Flying debris adds urgency: {snow spray|rain splashes|falling leaves|dust and grit}.
Rendered as a {photo|painting|clay figures|digital illustration|cartoon|oil on canvas}, with {dynamic motion|predator-prey tension|emotional contrast|cinematic framing|playful chaos} and {cinematic lighting|soft natural light|dramatic shadows|warm glow|diffused daylight|golden hour tones}. Shot from a {low angle|eye level|tracking side view} with {motion blur|shallow depth of field|panning streaks}.

Result:

[nabileisa]

Thank you for your hard work, couple of questions if you don't mind :

Is the fp8 quant model at https://huggingface.co/T5B/Z-Image-Turbo-FP8 supported?
Is there any difference between this huggingface.co/worstplayer/Z-Image_Qwen_3_4b_text_encoder_GGUF and the quants available at https://huggingface.co/unsloth/Qwen3-4B-GGUF

[NoMansPC]

Realism is astounding for such a small model. The team behind it really knew their target customer.

[250275196]

Server error: Given normalized_shape=[2560], expected input with shape [*, 2560], but got input of size[1, 256, 4096]

[jmguru]

Yeah I hit that error as well. I'm using a custom comfyui server, so I updated ComfyUI and comfyui_tooling_nodes on my custom server . I was able to render after that.

[250275196]

I still haven't been able to solve it.

[ETCore7]

In case you haven't tried yet:

kijai/ComfyUI-WanVideoWrapper#793

[jmguru]

Oh man, this model slaps even with all of my weird shit. It does horror and gore pretty well out of the box. I'm hearing there will be an edit version.

[ETCore7]

Here is explained how to get more variation across seeds with Z-Image Turbo.
https://www.reddit.com/r/StableDiffusion/comments/1p94z1y/get_more_variation_across_seeds_with_z_image_turbo/

[godlatero]

How to add Z image config in plugin?

[fukc-gihtub]

Open Preferences -> Styles, select Z-Image Turbo from the dropdown. If it says "checkpoint not installed", update ComfyUI and comfyui-tooling-nodes. Restart ComfyUI and Krita.

[icreaw-blip]

thank you,I successfully ran KRITA Z IMAGE using your method.

[HAMSAILY]

yes Controlnet-Union for zimage is here

https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union

https://civitai.com/models/2186383/z-image-turbo-fun-controlnet-union?modelVersionId=2461802

[Andree11gor]

релиз ещё не вышел.

если вы скачаете последнюю версию comfyui и последнюю версию плагина krita из git, то сможете их использовать

Please tell me what exactly needs to be updated, comfyUI or manager to the latest version? What version exactly do you have at the moment, that Krita AI has the ability to support controlnet for z-image?

[Vostdolom]

How can I update ComfyUI without reinstalling the entire KritaAI app? I indicated the version in the screenshot. I want to update everything to the latest version, but I don't know how to do it through the app itself. So, do I have to delete everything and reinstall it? Also, how do I add the "Manager" extension to Compfy, which is located inside KritaAI?

[fukc-gihtub]

How did you install it, in the first place?

[Vostdolom]

I downloaded KritaAI from GitHub and installed it last year. I hadn't used Comphy's internal nodes until last week, but for optimal performance with ZIT, I need a newer version of Comphy and, preferably, Manager to update new nodes.
I don't know how to update this "built-in" Krita Comphy now.

[fukc-gihtub]

I've never used the managed server myself, but from what I can tell, the plugin just asks for a folder to install ComfyUI into. You should be able to cd into that folder and update everything from there.

[xinru-liu]

Hi, so great to see the update!
What comfyui version is compatible with this new release? does it work with the latest comfyui - v0.3.76?

[nabileisa]

yes, z-image support was added to version v0.3.75 of comfyui with further improvements at v0.3.76 , check the comfyui release notes here https://github.com/comfyanonymous/ComfyUI/releases

[DenRakEiw]

thank you ^^

[Vostdolom]

Can you tell me how to update Comfy in Krita AI?
Will there be a Comfy update in the next Krita update?

[AlexYez]

ControlNet for Z Image v2.1 released: https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0

[jeroll3d]

Works fine, to who have a NASA computer

(:

[Acly]

1.44.0: added support for Z-Image-Turbo-Fun-Controlnet-Union

See top post for model downloads.

Can't set a start/end range for now, that will be added later. If the control-net causes too much image degradation reduce the strength or do a low-strength refine without control-net afterwards.

[nabileisa]

Hey @Acly, You might want to update the post since they just dropped an new and recommended 8step variant of the version 2.1 of both the tiled and union controlnets with claimed better quality and clarity at distilled speed.

[Acly]

Yea I was just testing it.

Compared to the non-distilled version it no longer looks blury and washed out, but IMO it tends to overshoot a bit into the other direction. It looks a bit fried like what you get when you run a 8-step model with too many steps.

[nabileisa]

Oh, that's a bummer , so can that be remedied by reducing the controlnet strength? , or do we still need a way to set a step range to prevent that from happening?

[agoulddesign]

Ran across this workflow for direct 4K generations. Didn't dig too deep but hope it helps with higher res gens and using the model as a refiner

https://huggingface.co/dx8152/Qwen-Edit-2509-workflow/blob/main/new-Z-Image-%E7%9B%B4%E5%87%BA4K.json

(Ignore qwen)

[jeroll3d]

I ll wait a better version, here i had problems with this model (nvidia RTX 3060 - 12GB)

[frikiti]

Please Acly,
"EulerDiscreteScheduler" is a MUST, to use Z-iMAGE or the render images get blurry and noisy.
¿How to manually add the sampler scheduler??
https://github.com/erosDiffusion/ComfyUI-EulerDiscreteScheduler
https://docs.interstice.cloud/samplers/

[christianv78]

I downloaded the exact text encoder you recommended (not the GGUF one) and put it in the appropriate folder, yet Krita is still telling me that it's not installed. Anyone able to assist?

[christianv78]

It seems that my issue was that I had put the file into the ai_diffusion/server/models/text_encoders folder and NOT the ai_diffusion/server/ComfyUI/models/text_encoders folder. It actually drives me nuts that both of these exist and I never know which is the correct one.

[Acly]

ai_diffusion/server/models/text_encoders is correct and works for me

don't put it into ComfyUI folder, it also works, but will prevent updates (which replace ComfyUI with new version)

[christianv78]

Good to know, but it did not work for me until I put the models there. I dunno why, but it is working now and would not before I moved them.

[fvladimir2005-oss]

I can't install controlnet. I download the file to the model_patches folder, but the control layers still don't work. Writes: "Not supported for Z-image". Can you tell me what to do? All files are located in the Frita - ComfyUI folder.

[HurricAngel]

I am having the exact same issue. Just to mess around I chose to upload the Fun-Controlnet safetensors file as lora for Z-Image and set it to around 20% it actually changes the inpaint outcome. Still I was not able to implement control layers but I was just feeling adventurous 🤣

[jmguru]

I've been experimenting with the new update of Turbo Fun Union controlnet Z-Image-Turbo-Fun-Controlnet-Union-2.1-2601-8steps.safetensors and it's a huge improvement IMHO. I don't need to run a refine after rendering using canny or depth now.

[fvladimir2005-oss]

I was able to solve the problem with installing controlnet, but now, when generating an image, I see these errors. What is the reason for this? I will also attach the specifications of my laptop. video card: NVIDIA GeForce RTX 4080 Laptop GPU, 12GB

[jmguru]

It looks to me that there's something wrong with the backend. Did you look at the client.log or server.log? Is this local environment or remote ComfyUI server?

[fvladimir2005-oss]

I looked at it, but I don't understand programming very well...I don't know what exactly to look for there. Moreover, Krita itself pre-installs and controls everything necessary. As for ComfyUI, I used a local Krita server, but now I have installed a portable ComfyUI separately to work through a configurable server in order to expand the possibilities of interaction. I haven't checked yet whether the full Z-Image-Turbo model works via portable ComfyUI. But I found a temporary solution for now by downloading the GGFU version of Z-Image-Turbo.

[Acly]

Updated the Links for ControlNet.

Strongly recommend to delete old controlnet models, and download the latest "2601" versions from https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.1/tree/main

The "Lite" variants are great too (small, fast, quality is perfectly fine).

"Tile" controlnet support was added in plugin version 1.46.0

[jmguru]

Thanks for the update. On Mac M4, I'm trying to use the lite variant but I'm hitting this error:

size mismatch for control_all_x_embedder.2-1.weight: copying a param with shape torch.Size([3840, 132]) from checkpoint, the shape in current model is torch.Size([3840, 64]).

FYI - I've updated the plugin, tooling and inpaint nodes and both 2601 control nets are both in model_patches What am i missing? TIA

[Acly]

needs comfyui update too

[jmguru]

I'm one of those ComfyUI Desktop on Mac weirdos. I think Desktop lags behind ComfyUI versions, so I'll have to wait. But non-lite versions work great.

[R4290N]

Why can't I use the ZIT models for FP8 that I download from Civiai?

Krita AI Diffusion:
Server error: ERROR: clip input is invalid: None
If the clip is from a checkpoint loader node, your checkpoint does not contain a valid clip or text encoder model.

How can I fix this? I have 8 GB of VRAM.

[jmguru]

Any news or rumors on the street regarding z-image edit ? I've been checking huggingface and google from time to time. Was it just a dream? However Klein9b is amazing.