Post "Z-IMAGE TURBO THE NEW AI KING! < 8GB VRAM REQUIRED! ABSOLUTELY INSANE!" by "Aitrepreneur" from patreon

Jesus Nuesi

Hi, there is an error on workflows, it seems some custom nodes where updated by the author and now the workflow does nto work, for example FastLaplacianSharpen FastFilmGrain, I have been installed manually but still do not recognise.

Starling

I had the same problem. So I replaced the "Fast Laplacian Sharpen" node with "ImageSharpen", with sigma and alpha set at .5. ImageSharpen is built in to ComfyUI and it seems to work just great.

傑生尚

Hi K. Thanks for the great workflows! My zit images have all of a sudden started coming out less crisp and sharp than before. I’ve tried tweaking a few things including sharpness and grain but no difference. I also deleted and re-downloaded the workflow to make sure I kept your default settings but still way less quality than before. It’s happening both locally and on runpod. Any idea why?

乃ㄩ$卄乃ㄩ!ㄥᗪ丂

this one click thing dont work , and when i try update comfui i get the below stashing current changes nothing to stash creating backup branch: backup_branch_2026-01-19_08_23_17 checking out master branch pulling. Traceback (most recent call last): File "C:\Windows\System32\ComfyUI_windows_portable\update\update.py", line 67, in ref = repo.lookup_reference('refs/remotes/origin/master') KeyError: 'refs/remotes/origin/master' During handling of the above exception, another exception occurred: Traceback (most recent call last): File "C:\Windows\System32\ComfyUI_windows_portable\update\update.py", line 70, in pull(repo) ~~~~^^^^^^ File "C:\Windows\System32\ComfyUI_windows_portable\update\update.py", line 11, in pull remote.fetch() ~~~~~~~~~~~~^^ File "C:\Windows\System32\ComfyUI_windows_portable\python_embeded\Lib\site-packages\pygit2\remotes.py", line 245, in fetch payload.check_error(err) ~~~~~~~~~~~~~~~~~~~^^^^^ File "C:\Windows\System32\ComfyUI_windows_portable\python_embeded\Lib\site-packages\pygit2\callbacks.py", line 111, in check_error check_error(error_code) ~~~~~~~~~~~^^^^^^^^^^^^ File "C:\Windows\System32\ComfyUI_windows_portable\python_embeded\Lib\site-packages\pygit2\errors.py", line 67, in check_error raise GitError(message) _pygit2.GitError: failed to create temporary file 'C:/Windows/System32/ComfyUI_windows_portable/ComfyUI/.git/objects/pack/pack_git2_c791ee8fdd9a2042': Access is denied. Press any key to continue . . .

kaczorefx

Hi, the detail quality is amazing but the overall image looks like it's been compressed somewhere during the workflow. I keep seeing banding on flat patches of color - especially on walls. Instead of a smooth white/beige gradients I see bands of various colors, magenta, green, cyan. Look a little like a Gif saved with 256 colors. Any idea what could be causing this? It's especially visible in images generated with the Detail Demon workflow.

Joshua Lebrun

solved the issue by matching cuda version in python embed folder with installed cuda version on pc, ended up using 12.6 which works on gtx 1080. However gen time is around 6 minutes on the included workflow which is wayyy to slow.

Joshua Lebrun

Hello K! I've been a big lurker on your videos for quite some and left the AI space when auto1111 was the main way to work. Now I'm back and trying my hand at the z-image workflow on comfy. However following the install instructions I get this error when trying to generate an image: "CLIPTextEncode CUDA error: no kernel image is available for execution on the device Search for `cudaErrorNoKernelImageForDevice' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions." Could you help me out with this please? I am running a 1080gtx. Thanks!

Mark Sykes

I have solved by swapping to the lora loader stacker blew my mind for a good minute

Mark Sykes

This really is awesome but I have hit one major snag with the lora loader. The add lora button is stuck right in front of the lora weight slider. For ages I thought I was clicking on the add lora button and swapping it and something was cooking my images but it was stacking them and I cant change it or view the lora to adjust weight… its pretty frustrating is there any fix?

Rodger Hurlburt

This is the most badass set of workflows I've seen, the lora I made works quite well with the controlnets. I put together a psuedo controlnet myself but it didn't work nearly as well as this. its like hugely better than the one I janked together ahahah

Aitrepreneur

Yes if you downloaded the comfyui amd version

Han

Can this be run on AMD GPUs (RDNA4)?

Capt Jebus

Don't worry, Solved it, I had accidently turned on Nodes 2.0 Beta, that caused the issue of compacting the the loras.

Capt Jebus

Does anyone know how to fix the lora loader where the loras are compacted and cant be seen correctly picture of problem in the link: https://postimg.cc/5j0kTnBm

Aitrepreneur

I'll update the workflow and make a quick video once I'm done testing it

Ken L. Black

I just saw that version 2 of the Controlnet is out, but it appears to only be the full version and not the fp8 version like you have included in the workflow. Has anyone out there tried out version 2 in the "Psuedo Controlnet" workflow and got it to work?

Aitrepreneur

I always make the manual install guide with the link to models

Martim M Costa

Big fan of yours and hope you keep up the good work. You are officially my go to when it comes to all things Gen AI (at least running locally). Minor constructive feedback: I understand you have the 1 click installers but for those of us manually installing the models and extensions it would be useful to have the links to download the models, clips and VAEs. as some of those are not evident.

Aitrepreneur

Oh wth...I didn't notice that! I mistakenly used the preview image node instead of save image node! Thanks for the heads up! I've updated the workflow you can just redownload it from the post: https://www.patreon.com/posts/144536359

Ken L. Black

First of all....FANTASTIC workflow! I've really had a lot of fun with it. Being a bit of a noob at ComfyUI, I wonder why--unlike the other workflows--that the ControlNet workflow does not save the images generated automatically. It's no issue to just right-click and SAVE either the temp one or the final one, but I'm just curious. Is there an easy-(ish) fix for this so that they'll just save automatically?

Aitrepreneur

if you take an image that was generated inside comfyui you can just drag and drop it inside comfyui (like a workflow) and it will give you the workflow + precise prompt used to generate that image

Aitrepreneur

dm me the full error log

JJ Platz

Hi, i updatet Drivers and Comfy: But there is this Error: ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\lib\c10.dll" or one of its dependencies. Can you help, please?

Stefan Meister

My Dear Overlord. In one of your workflows you had a simple solution for extracting a prompt from a given image which unfortunately did not work for me. So could you please, if you can, point me in the direction of a simple but well working solution for this. I've googled, but most of the "solutions" were no solutions at all, or, if lucky, some strange overbloated show off workflows.

ElecMat

All good now !Upgrading my Nvidia drivers to the latest solve the problem!

ElecMat

I tried the installation locally using the .bat file but there's an error message during the installation process that says "ERROR: Could not find a version that satisfies the requirement mediapipe (from versions: none) ERROR: No matching distribution found for mediapipe Then the installer ask me to press a key to exit. I'm unable to start ComfyUI! Any idea what is going on ? Thx

Ben Holmes

I finally got a chance to get a runpod loaded but when I loaded the v2 of the z-image turbo workflow below the textbox for entering the prompt and negative prompt show up so vertically thin I can't read or enter anything. I've tried everything I can to get it to expand and haven't been able to get it to work. In addition, the regular manager is now missing...

Stefan Djokic

Nice one! Btw New Controlnet model for ZIT is out v2 last night, I tried to manually change the modelpatchloader node to it but seems I can't (was a full not scaled modeL) Are you going to update it again? :D

Teresa Greasley

On the detail deamon the 2 images were generated but did not show up on the Image Comparer Node, am I missing a setting?

Bedroom Guides

How to download on MAC?

LW

thanks, will give this a try

Alex Roca Sánchez

Protip to get different faces using different seeds and better details: use in the KSampler node dpmpp_sde as sampler and ddim_uniform as scheduler

Dirk

Ok, solved it. Just in case anyone else get this error: just do an comfyui-update via the bat file :-)

LW

Hi K, thanks for the new ZIT workflow (OMG.. they should have thought of asking a native English speaker before calling the model Z-Image-Turbo) with all the new cool features. One question though:- In your video showing the Viking girl, it seems that whatever the seed, the Viking girl has the same face. Even when using the Variance node, the girl has essentially the same face. The pseudo controlnet also does not overcome this issue. Your lovely green-leaf girl suddenly changes to the Viking girl's face. I have noticed this kind of face fixation with other models and I was wondering if this is simply a limitation of distilled models? Maybe a distilled model thinks all Viking girls look the same? I know that you can influence the face (somewhat) with the prompt but I would have expected the seed to have a much stronger effect on the face, as it often does on other parts of the image. Any thoughts on this? Or suggestions how to get a bit more variation?

Jaina Shop

THIS IS THE BEST WORKFLOW Z-TURBO! I've been trash my other folders contain : Flux, WAN, and QWEN, that are quickly surpassed by this great model! I can upscale with extremely good (God) quality easily at 9000px in high. just FANTASTIC! The updated V2 workflow works GREAT! ... Now I can't wait to see come out for the Base version and the edit. I'm using the non GGUF, quality is better. I'm creating images at 2248x1536px with incredible quality, took for my 3090TI about 30-40 seconds at this size. Pro tip: push the pedal on the ModelSamplingAuraFlow :)

Gavin

Hello, I'm new to this. For img2imag, how do you generation more than one image at a time?

Dirk

Thanks for your work! I really love the Z-Image! The V2 Workflow runs smoothly for generating images, but when I try the Controlnet, I get an error "cannot access local variable 'model' where it is not associated with a value". The Z-Image-Turbo-Fun-Controlnet-Union-fp8-e5m2.safetensors file is in place, but ComfyUI geives me an error at this node. Any hints? Thanks!

Luis Muniz

it's in this post: https://www.patreon.com/posts/144536359

Luis Muniz

Where's the workflow file?

Zolon Farsane

Can you make a full installer I can use on ubuntu? I have 2 5060ti 16gb cards on that system.

Alex Roca Sánchez

For people with Nvidia GPUs of Pascal or Maxwell achitectures (GTX 900 series, GTX 1000 series...) like me, this won't work but it can be fixed. The issue is because installer uses a Pytorch 2.9 compiled with cu130. Since Pytorch 2.8 the Pascal and Maxwell support continues only in the Pytorch versions compiled with cu126: https://dev-discuss.pytorch.org/t/cuda-toolkit-version-and-architecture-support-update-maxwell-and-pascal-architecture-support-removed-in-cuda-12-8-and-12-9-builds/3128 Meaning, after installing the .bat, users with these GPUs have to go to the folder with the run_nvidia_gpu.bat file and write this in CMD to reinstall a current PyTorch that has Pascal & Maxwell implemented: python_embeded\python.exe -m pip uninstall torch torchvision torchaudio torchdata torchtext sentencepiece -y python_embeded\python.exe -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

Aitrepreneur

Sure. here: https://we.tl/t-JNqkpMdr43

Alex Spin

Can you make a NON GGUF version of this? ... using the normal high quality for high mem GPUs? like from 16 gb and above?. I've notice that the non GGUF it has a little more details and overall quality.

Aitrepreneur

bitdefender is known for that...hence why I don't use any third party antivirus softwares. just disable it before running the bat file

Aitrepreneur

model_patches

Aitrepreneur

if you used the runpod installer, you don't need to update comfyui with the installer, just updating it through the manager (update all button) is enough

Can Turan

probably a false-positive but bitdefender quarantined the UI\Z-IMAGE-TURBO-MODELS-NODES_INSTALL-V2.bat. It was ok with v1

Matthijs

What"s the folder for the controlnet model control net ? THX.

Thug Tear

When I go to install custom-nodes, it just spins endlessly and doesn't load the ones needed. Also why can't I "Update ComfyUI" when using a storage? It just says failed to update comfyui. I run "chmod +x Z-IMAGE-TURBO-AUTO_INSTALL-RUNPOD.sh ./Z-IMAGE-TURBO-AUTO_INSTALL-RUNPOD.sh" ComfyUI update fail: Cmd('git') failed due to: exit code(1)

Aitrepreneur

what do you mean by not updated? dm me screens and error logs

Aitrepreneur

DGX spark?? Jesus man how much did you pay for it? I'm jealous😆 It literally just came out so I haven’t tested it myself yet, but since it's Blackwell-based it should run everything fine. The only problem is that it's Linux based so most of the tools I showcase in my videos are in windows but everything should work on Linux as well, just that following along with my videos will be a bit harder for you but otherwise it's cool

NeverEnoughHats

It keeps saying during my clean install that ComfyUI isn't updated and I updated it using the bat file. I don't know how to troubleshoot this.

rbr194

I have been following your videos a long time and Im a big fan. I bit the bullet and just bought a DGX spark to force myself to dive deeper into AI and learn more on the development side. I have been doing most of it in Windows so it will be a bit new to me. How do you think it will handle with a lot of the image generation models and stuff? I currently have 2 tesla P40 GPUs in my desktop but find they are outdated for some of the stuff you post.

Aitrepreneur

Nope, it's in its own virtual env

Aitrepreneur

I'll be making a quick update video about it. Euler simple is great, dpmpp_sde and ddim uniform is maybe even better

Aitrepreneur

it's in this post: https://www.patreon.com/posts/144536359

Aitrepreneur

It's a false positive, happens so often with 3d party antivirus softwares like bitdefender.... Disable it and then run the bat file

Alessa

Hey will the AI-toolkit wan 2.2 lora 1-click installer mess with my current installed packets? I dont want to break forge, comfy, etc

Paul Mazurkiewicz

Just curious, what sampling method and schedule type are folks using for best results?

Eager Alice May

strange..... installed the nodes installer, but no json workflow....

Tom Baron

My BitDefender antivirus went crazy when I tried running the Z-Image-turbo-models-nodes_install.bat all sorts of my executables were trying to be swapped and were quarantined by bitdefender. Has this happened with anyone else here?

Aitrepreneur

That's not what image to image does, it's not an edit model. Image to image changes the entier image depending on the prompt. Use the Qwen image edit 2509 workflow for specific changes (or the qwen image edit 2509 inpainting workflow)

Aitrepreneur

It's funny how pretty much none of the things you just said is correct, I don't even know where to start here... 1) Women don't look "samey" if you prompt correctly, you could technically say that for any model released because of bias training for every model 2) Not sure what you mean with the patchy skin, probably the z-image specific noise applied but even then, calling it skin disease is so weird or the model is not working correctly on your side 3) Not faster than Flux or WAN?? is just flat-out wrong. Again, not sure if it's because the model is running incorrectly on your side but it's so much faster than Wan it's crazy x2-x3 faster for me, all the while using less than 2x the vram... 4) And saying it’s not a leap forward… dude, this thing is one of the biggest upgrades we’ve had in a long time. The amount of details and precision this model brings in such a small package has never been seen before and it's still not the base model yet, just the distilled turbo model...hence why the community is crazy about this model right now... If you don’t like the model, that’s fine, but almost everything you said is just wrong. If it’s installed and running properly, the improvement over ai other models is obvious.

Paul Mazurkiewicz

I agree with the comments that all the women look very samey, kind of this Chinese chick kind of look. Also, their skin is really patchy, like they have a disease. This model is meh really, there are much better out there. Its not any faster than Flux or Wan either. Maybe there's some promise with fine tunes and loras I guess. It does run quite well on low memory systems, but I wouldn't call it any great leap forward really. Just another attempt at China to show everyone they are in the game with limited compute resources. Good for them, competition is a good thing.

Raja Segar

Guys need some help, Text2Image works fine, however image 2 image generates the exact same image, just very slight changes from the source image. I asked it to change the color of the clothes, nothing happens but the face changes slightly. I did not change anything from the default workflow. Denoise set to 0.3 does nothing also. Please help. Thanks

a pretty parrot

You should be able to use the runpod one. Just put it in your ComfyUI dir and chmod +x FILE and then run it with ./FILE.

a pretty parrot

All of these files? They are designed for different os, you should never need more than one. Theres nothing really special about it that is truly runpod specific. I made my own set of scripts to kickstart a runpod to do the same thing

a pretty parrot

You probably can, but quality probably won’t be great. The edit model is what you’d really want for that. I am eagerly waiting for it as well

PoliticHoles USA

Do yoiu think that you will take alll of these and combine into 1 zip file?

Hrothgar's Hoard

I like this and have added an auto prompt workflow so I can upload a picture and get a good prompt from it. Is it possible to do a face-swap in imagetoimage or inpainting?

mnkjiu

Do not update comfyui at the moment. if you are encountering a need to update nvidia drivers, copy from a working copy of your comyfui portable python embedded folder and replace the broken comfyui folder.

Kink Alchemist

Any script for local linux installation?

Klaus The.

Got the same with the Q8_0 model. With the Q6_K it is working, generating images on a 4090 in only a few seconds.

UnderToker

Getting this error when attemping to run... size mismatch for model.layers.0.input_layernorm.weight: copying a param with shape torch.Size([2560]) from checkpoint, the shape in current model is torch.Size([4096]). size mismatch for model.layers.0.post_attention_layernorm.weight: copying a param with shape torch.Size([2560]) from checkpoint, the shape in current model is torch.Size([4096]). size mismatch for model.layers.1.input_layernorm.weight: copying a param with shape torch.Size([2560]) from checkpoint, the shape in current model is torch.Size([4096]). size mismatch for model.layers.1.post_attention_layernorm.weight: copying a param with shape torch.Size([2560]) from checkpoint, the shape in current model is torch.Size([4096])

Jason

Using the Zturbo... worked fine for a couple days and now all the checkpoints it says are missing in the UI.... how do I get them back?

jenson clarke

289 seconds to load the picture with q5 on a 2070super :(

Aitrepreneur

my installers are for windows, since it's what I and most people use, I don't do specific linux installer because I can't make sure it works on my side but if you have comfyui installed maybe the runpod installer should work immediately

Aitrepreneur

Are you using my workflow here? I don't use those nodes in this workflow

Aitrepreneur

video/gif? are you talking about the correct workflow here? Are you talking about the other expressions workflow? Look in your output folder for the video

Zolon Farsane

I have just moved my dual 5060ti 16gb cards to a headless ubuntu sever. Still muddling my way through it, and I was wondering how I can install this on the system? Thanks.

Solid Snake

wmic is legacy sys info tool from win 10, removed for 11, it's called when attempting to detect your GPU. it's not there so it was flagged, but if the installed completed you're good. you can see it's defaulting to PyPI (python package repo). did the installer complete?

Solid Snake

The loader was designed for standard CLIP architectures (OpenAI CLIP, etc). it doesn't support Qwen3-based text encoders yet my G

Solid Snake

well done mate!

Magnus Nelénius

Hi, I tried the Z-IMAGE-TURBO-MODELS-NODES_INSTALL.bat and got: Which Z IMAGE TURBO quantization? 1) Q5_K_S – GPUs < 8 GB VRAM RECOMMENDED 2) Q6_K – GPUs 8-12 GB 3) Q8_0 – Best quality, GPUs 12-16 GB and more Enter 1, 2 or 3: 3 'wmic' is not recognized as an internal or external command, operable program or batch file. No local wheel dir found. Will use PyPI. What is wmic and how to fix it?

Solid Snake

yo the guy's health has been suffering. we're a community too - so what's your issue G maybe we can assist?

Solid Snake

yeah this will happen when you use AI to install your AI.... which is what I had been doing. two dif versions of py and a few installs of comfyui somehow. The .bat builds the folder wherever you run it too, so it took me a little bit of time before I realised the comfy dir was in there ... silly me.

Danilo Hugo Arreghini

hi, im having the following issue: CLIPLoaderGGUF Unexpected text model architecture type in GGUF file: 'qwen3' im using clip "Qwen3-4B-UD-Q6_K_XL.gguf"

Yukon Cornelius

to anyone having issues. I cannot guarantee this will work, but its a good idea anyway. I use a program once in awhile called "WinDirStat" which rapidly scans your hard drive and then gives you a visual map of what is on your hard drive. I realized that I had a RIDICULOUS amount of stuff related to AI that I was not really using. I am talking almost a TB of crap. Like.. I literally had 5 types of python installed that had accumulated huge amounts of space and. well if you run it on your machine you will see what I mean. It is also just good to run once in awhile and I swear you will find that half of the data you have on your hard drive is just old crap you aren't even using. It is a wonderful tool to play around with and see all the stuff you have forgotten in folders you forgot even existed deep inside your drive. The reason I say it may work for this though is that I noticed a lot of the issues I would run into is because comfy ui can get really confused about what version of python to use, or what wheels or blah blah blah torch this or that. its usually because you got half a dozen different versions installed at the same time. just go through your system with that. nuke everything from orbit until its all gone. and just start over. It doesn't take very long and everything really runs pretty smooth if you stay on top of this sorta thing.

Nmuster093

I used the one click installer. Error by start of the run_nvidia_gpu.bat workflow: “ This workflow uses custom nodes you haven't installed yet. DetailDaemonSamplerNode LyingSigmaSampler “ By running the workflow with the activated „ Prompt execution failed Cannot execute because a node is missing the class_type property.: Node ID '#358' ”

Kaiser Lowendachs

I have found that the 1 click installers generally work, at least when they are released. I think the old stable diffusion and flux workflows have broken somewhere, I think because nodes are no longer available in comfyui and the updates break them. It might be worth going back and looking at some of them and seeing if they can be updated - though no sure how much work that would be.

Jeb

Installed on runpod as per directions. Did not produce the video/gif. No errors or anything, just never put out the gif. Last thing it said was "prompt executed in 14:13". Great idea, looks good in the video, doesn't work.

Manu

I need to be honest about the one-click installers because it’s not okay anymore. Out of 10 one-click installers, maybe 2 worked without me doing anything. The rest only work after I spend hours fixing them with GPT, Gemini and every trick possible. That’s not normal. You advertise them as ONE-CLICK installers — that’s literally why people pay — but in reality 90% of them don’t work straight out of the box. And yes, I installed everything exactly how you show in the videos. They don’t work anymore because the node repositories and other dependencies have changed, and nobody is updating the installers. This really needs to be addressed.

Manu

Hey, I really need to say this because it’s getting too much. I’m paying £54 every month for the Royal Emperor tier, and honestly… I don’t see the point anymore if the support is like this. I sent you a message on October 9, you replied on October 18, and since then nothing for almost two months. And even that reply was basically nothing helpful. Why am I paying for Royal Emperor if I get one short answer every 2–3 months? It just doesn’t make sense. I’ve been supporting the project because I believe in it, but being ignored like this really isn’t okay. Please check my message and actually help me out.

Aitrepreneur

that's not normal, can you dm me screenshots of the parameters and full error log

Aitrepreneur

We have so many models coming soon though, not sure which one will be on top.. Qwen image edit 2511, Z-image base, Z-image edit.... I didn't put the bf16 version up since for most people they'll want to use the quantized versions, only use bf16 if you got more than 16gb of vram for the full speed

Aitrepreneur

Not sure either, they all work fine

Aitrepreneur

what's the issue?

Aitrepreneur

it's the https://github.com/vrgamegirl19/comfyui-vrgamedevgirl node, it should be installed, if you already had it installed before, update it from the manager

Aitrepreneur

Because you ran the installers as admin I suppose, don't do that, just the bat file normally

Aitrepreneur

Man what...? Seriously? If you have any issues during install I'll be happy to help. But if you don't even say what issue you have I can't help you fix them

Aalon Mayers

feeling like this guy robbed me, none of these so called one click installers work. they just get you half way then you have to figure out the hard part. so why did i pay to join your patreon

FherraZ

Hi, I would like to ask for some help, I'm getting this error: CLIPTextEncode CUDA error: no kernel image is available for execution on the device. UPDATE - FIXED If anyone has same problem, do this in powershell inside folder python_embeded .\python.exe -m pip uninstall torch .\python.exe -s -m pip install typing_extensions==4.15.0 torch==2.7.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126 I use cu126 because my gfx is old, I hope it helps.

Steven Melgar

Hello, I'm having a problem updating comfyui. This is the error I get: Traceback (most recent call last): File "C:\Windows\System32\ComfyUI_windows_portable\update\update.py", line 70, in pull(repo) ~~~~^^^^^^ File "C:\Windows\System32\ComfyUI_windows_portable\update\update.py", line 11, in pull remote.fetch() ~~~~~~~~~~~~^^ File "C:\Windows\System32\ComfyUI_windows_portable\python_embeded\Lib\site-packages\pygit2\remotes.py", line 245, in fetch payload.check_error(err) ~~~~~~~~~~~~~~~~~~~^^^^^ File "C:\Windows\System32\ComfyUI_windows_portable\python_embeded\Lib\site-packages\pygit2\callbacks.py", line 111, in check_error check_error(error_code) ~~~~~~~~~~~^^^^^^^^^^^^ File "C:\Windows\System32\ComfyUI_windows_portable\python_embeded\Lib\site-packages\pygit2\errors.py", line 67, in check_error raise GitError(message) _pygit2.GitError: failed to create temporary file 'C:/Windows/System32/ComfyUI_windows_portable/ComfyUI/.git/objects/pack/pack_git2_a141b2e79c4062b1': Access is denied Any help please?

Jordan Baker

having trouble with the "FastLaplacianSharpen" and "FastFlimGrain" nodes. What node pack are they from? Not showing up as missing in the manager.

Stefan Djokic

Lol only Z-Image loras. This model just came out so there's not a lot of them but you can check out on Civitai. Anyway there's also a way to train them now so we'll see soon

YoungMoneyMaker

can someone help me connect to the run pod pls, im having troubles

Philip Smith

what loras are compatible with this model?

Yukon Cornelius

not sure why everyone is having issues with qwen after this.. i just installed the nodes, which downloaded the models. and then updated comfy ui and.. tested both this and qwen, worked fine.

Yukon Cornelius

Pretty Sweat. I think for me I will be sticking with QWEN until they release the "edit" model. really good though. I also noticed there is a bf16 version that might be worth checking out if you have a bit more vram. For some reason it generates images really easy and fast for me, but when i do inpainting it was saying it was going to take like 14 minutes for one image, and it wasn't like some giant image. kinda strange.

Nick Enlowe

Lazarus pit

FINALLY a high quality simple functioning process this thing is insane

Aitrepreneur

Yeah just windows stuff, it happens. When you generate images, it uses your C drive as temporary virtual mem, so it look like you lost space and when you close the UI, those temporary space allocations get cleared and all the space comes back.

Aitrepreneur

How so? It still works for me, nothing changed

Aitrepreneur

Well… yeah, because you’re comparing apples to oranges here. Illustrious is a special anime fine-tune of SDXL, around 2.6B parameters, while Z-Image Turbo is a 6B general model. Of course Illustrious will use less VRAM and be faster on your 3090. Saying that Z-Image Turbo “doesn’t respond well to prompts” is only true if you’re trying to do anime and comparing it directly to an anime-specific model. Illustrious is literally trained for that niche. It’s like saying “this knife cuts bread better than my phone, so the phone sucks.” The knife was made for that exact job, the phone wasn’t. If we fine-tune Z-Image Turbo on the same dataset as Illustrious, it should perform even better in that domain. The team already mentioned they want to release a dedicated anime fine-tune too, so we’ll see when that one drops.

BopDobop

This works but it's not fast at all. I run it at 3090rtx and this works slower than sd.next or WebUI Forge with illustrious model. For example WebUI Forge creates 768x1024 px prompt with illustrious model within 5-8 sec, using 7 gb vram, while this needs up to the 15 secs (I turned filters off) and up to the 12 GB vram using the same prompt. Also loaded model doesn't respond well to prompts

Shelby Black

Qwen image edit stoped working.

creep

ehm... ok... i've been using local ai for some while now but never encountered this .... when i run this and i generated images it bites a huge chunk of my (C:) disk space ... like i literally had ,lets say, 50gb space and i generated 1 image and i had 30gb... so i generated another one and i had 10 and so on ... untill i hit 0.... then i close the ui/bat and all the GBs are back to normal ?.... so what is happening :O ?? ( hopefully you understand me ). EDit: ok so now i tried to generate 10 images, starting with 110 gb and ended up with 115gb... xD

Klaus The.

Tip, for the ""AssertionError: Torch not compiled with CUDA enabled" - I had to uninstall with a ./ before the python.exe. Otherwise it just uninstall in the system-wide Python path (c:\users\...\appdata\local\programs\python\python310\lib\site-packages) but not in the "stand-alone" env for this project. So open a terminal/power shell in .../ComfyUI_windows_portable\python_embeded and then: ./python.exe -m pip uninstall torch ./python.exe -s -m pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

Aitrepreneur

yes, are you sure you are entering the width and height in the correct place?

Martin Gee

Hey, is it possible to create vertical images with this workflow? Even if I choose greater height than width the outcome will be a horizontal image every time.

LW

Yes, that was it. I downloaded the file from Hugging face and it is 7GB. Works fine now. Very fast!

LW

Hi K, I'm getting the following error: UnetLoaderGGUF cannot reshape array of size 5843709 in to shape (10240,480) I have updated the Nvidia driver for my 3090 (v32.0.15.8180) and updated comfyui (v0.3.75) Any suggestions? Just noticed that file z_image_turbo-Q8_0.gguf is only 971,886 KB .... that is surely too small...?

Aitrepreneur

you are not using the correct workflow...nowhere it uses the ae.sft vae...

Aitrepreneur

Update comfyui and update all the nodes using the manager, they are not up to date for you

Derek Dillman

I'm new to Comfyui. When I launch it, then just try clicking Run, I get an error (abbreviated below). I tried Updating everything.How to resolve? Prompt outputs failed validation: VAELoader: Value not in list: vae_name: 'ae.sft' not in safetensors Value not in list: clip_name2 clip_name1 UNETLoader: value not in list: unet_name

George Poulopoulos

i got that error: CLIPLoaderGGUF Unexpected text model architecture type in GGUF file: 'qwen3'

Aitrepreneur

It's a bat file my dude, you can edit it and look what's inside...."Remove my post lol, what is this some blackmail?🤣"

Lord Braska

Charles Leonhart

this is amazing

Aitrepreneur

You need to AMD comfyui portable version for it but it should yeah

Aitrepreneur

what's the issue?

Aitrepreneur

The node you need is: https://github.com/vrgamegirl19/comfyui-vrgamedevgirl but it should have been downloaded. make sure you have updated comfyui with the comfyui update bat file and you have updated all the nodes via the manager (and first make sure you got the node present as well

Aitrepreneur

I modified the installer and uploaded a v2 version, try this one

Aitrepreneur

Ok, I modified the installer and uploaded a v2 version, try this one

Brian Reyes

Hey, I wanted to let you know there is an issue with the auto install script. When it tries to download Git, it is not grabbing the actual installer. It ends up saving an HTML page as an EXE, and Windows throws the Unsupported 16 Bit Application error when I try to run it. Just letting you know so you can update the link or the script whenever you get the chance. I can test it again once you fix it. Thanks.

Sven van rompuy

When i run your batch auto installer I get a error message saying that the git-setup is incompatible with 64bit version and stops. how do i get around this ?

Tony White

I already have comfyui installedso I ran the Models-node-install.bat file in my comfyui folder. I started comfyui after that but it says that these 3 nodes could not be found. I installed missing nodes in Manager and updated everything but I still got the error message. Here are the 3 missing nodes. FastLaplacianSharpen Upscale by Factor with Model (WLSH) FastFilmGrain. I even did a search but they're not in the comfyui library.

Jason Nehl

Every since i built a new pc, I can never get it to work.

mnkjiu

Tested on Macbook Pro. No issues. Update ComfyUI. Install all missing nodes. Generation takes about 258 seconds for initial image generation and subsequent images takes about the same time.

Michael Bryant

Works like a charm. Thanks!

gregoryallen caldwell

will it go on a AMD integrated graphics machine with lots of ram and vram ?

gregoryallen caldwell

me too....... i have tons of ram and vram the way my pc is set up and modified but its STILL AMD integrated graphics, does it optimize CPU? i run sd 1.5 in olive optimized version fast and easy on it in my linix boot, was thinking of putting this on my win boot side.

ljbarrett5

which bat file should I run if I have an AMD GPU?

Aitrepreneur

you just need to wait for it to be ready, just be patient basically

Aitrepreneur

those installers are for windows, I guess they can be modified to work for linux but I can't make sure that they work on my side, you can just load the workflow and install the nodes through the manager though via the missing nodes tab

Aitrepreneur

HAPPY THANKSGIVING!

Jon

I was also getting Unexpected text model architecture type in gguf file "qwen3". Hitting update all in comfyui manager fixed it.

Thug Tear

did all this "Create an account if you haven't already: Runpod Click on Pod (on the left side) then click deploy Choose a GPU with at least 24gb of VRAM (4090 is best), click here for the comfyui template (aitrepreneur/comfyui), then edit the template and choose 80gb for both the container and volume disk, then deploy on demand but then I go to connect to port 3000 and I get this error "The port is not up yet NOTE: This is not necessarily an error Check your container logs for the following: Ensure that the applications have completed syncing to /workspace. Ensure that there are no errors. Ensure that the logs say Container is READY! If the container is READY Check your application logs (refer to the README for more information)."

CemeteryPumpkin

Same here. *edit - updated comfyui via the manager and it worked fine.

onawa

im getting unexpected model architecture type in GGUF file qwen 3

Xaro

Under Windows it works great but under Linux it always missing custom nodes and i cant Install/Fix them.

Kiw0x

Sorry guys, I must be missing something - ComfyUI works and everything but I'm missing a lot of the essential nodes. It states to download the 'Z-IMAGE-TURBO-MODELS-NODES_INSTALL.bat' but I can't find that specific link to the download? [edit] refreshed the page later on and the download links became available, I was probably too early or page bugged out. Thanks for the work on this AIT!

David C.V.

The error is; CLIPLoaderGGUF Error(s) in loading state_dict for LLama2: sine mismatch for model....(infinite list) Thx from Madrid!

David C.V.

First to all...THX like always. Issue with Z-Image Turbo GGUF: Qwen3-4B Size Mismatch (2560 vs 4096) Hi, I am trying to run the Z-Image Turbo workflow following your manual installation instructions. I am using the Qwen3-4B-UD-Q6_K_XL.gguf file for the Text Encoder as instructed. I have updated ComfyUI-GGUF to the latest version to support Qwen3, but I am getting a persistent RuntimeError: Error(s) in loading state_dict for Llama2. The error specifically states: size mismatch for model.layers.0.input_layernorm.weight: copying a param with shape torch.Size([2560]) from checkpoint, the shape in current model is torch.Size([4096]). It seems that selecting lumina2 (as in the original workflow) or qwen_image in the CLIPLoaderGGUF node forces a model dimension of 4096, but the Qwen3-4B file actually has a dimension of 2560. Could you please clarify which 'Type' needs to be selected in the updated CLIPLoaderGGUF node to support this specific Qwen3 4B architecture? Thanks!