Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

QWEN COMBO ULTRA WORKFLOW! (Patreon)

Published:
2025-08-11 20:04:17
Edited:
2025-08-11 23:47:23
Imported:
2026-06
Tags:

Downloads

Content

EDIT 12 AUGUST: 4 STEPS LORA IS OUT! THE WORKFLOW AND INSTALLERS HAVE BEEN UPDATED! SO NOW YOU CAN GENERATE AN IMAGE TWICE AS FAST...AI GOES REALLY FAST HUH..😅YOU CAN REDOWNLOAD THE WORKFLOW AND EITHER USE THE QWEN-ULTRA-MODELS-NODES_INSTALL.bat inside the comfyui folder or dl the lora from this link: https://huggingface.co/Aitrepreneur/FLX/resolve/e70ebc249a07b2076db848b166d8ff1c759ce953/Qwen-Image-Lightning-4steps-V1.0.safetensors?download=trueand put it inside the ComfyUI_windows_portable\ComfyUI\models\loras folderHAVE FUN! 🙃Hey everyone! I've just released a new video on the NEW QWEN IMAGE MODEL! THE NEW BEST TEXT TO IMAGE MODEL ever made!Insane artstyle, incredible prompt following, great text design and just amazing quality. It can do everythingl. BUT...NOT BY ITSELF...The only way to TRULY Make it the best is THE SUPER COMBO TEAM QWEN + WAN 2.2! And My Workflow makes this possible!You can check out the video right here: https://youtu.be/7P4LHEAEGNg

I also created a 1-click installer for the Webui for my Patreon supporters for Local AND Runpod, that you can find right here: https://www.patreon.com/posts/136261655

Here the link to the manual install guide: https://pastebin.com/18cv3Ww7

Down below you will also find the QWEN_COMBO_ULTRA_WORKFLOW.json for free for you to use!

As always, supporting me on Patreon allows me to keep creating helpful resources like this for the community. Thank you for your support - now go have some fun😉!

Files

Comments

Justin

It's always the best when new stuff comes out :)

Lance Fanger

I think Chroma is better over all currently

Brad

but is it as PG13 as Kontext?

s h a r k e y

A merge between Chroma and Big Lust would be the best for that stuff i reckon

Aitrepreneur

yes but loras are already here, wan was the same and look what we can do with it

Aitrepreneur

BEST ONLY when combined with wan 2.2, otherwise qwen is just the "best model at following the prompt"

Buraku Studios

What paid online portal is the best so far at combining the latest workflows and models?

Me Myself and I

Ideally I'd like a model that generates images indistinguishable from real photos. Unfortunately this is not it. Wan by itself is still better at it.

Aitrepreneur

Base model yes, wait until we get realistic loras...

Me Myself and I

Yes, at least there's a strong foundation with good prompt adherence for people to build on.

zoechelfyn

I can't find qwen_image_vae.safetensors in the main branch directory. Am I blind? can anyone help?

Aitrepreneur

vae file is in the models/vae folder if you used the installer correctly

Lance Fanger

Damn now I have to try a new model to see for sure... My poor hard drive

John Smith

I keep getting this error. Ive tried all three versions and they all fail. -------- Downloading ComfyUI -------- % Total % Received % Xferd Average Speed Time Time Time Current Dload Upload Total Spent Left Speed 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0 91 2028M 91 1850M 0 0 6277k 0 0:05:30 0:05:01 0:00:29 5427k curl: (18) end of response with 187232291 bytes missing Download failed. Press any key to continue . . .

Sidney Anderson

Pulling my hair out with this one: Prompt outputs failed validation: CLIPLoaderGGUF: - Value not in list: type: 'qwen_image' not in ['stable_diffusion', 'stable_cascade', 'sd3', 'stable_audio', 'mochi', 'ltxv', 'pixart', 'cosmos', 'lumina2', 'wan', 'hidream', 'chroma', 'ace', 'omnigen2'] Using the packaged workflow with the selected "Qwen2.5-VL-7B-Instruct-UD-Q4_K_S.gguf" Any help would be appreciated!

satsFlowMaxi

My images don't look as crisp as yours using the default settings and prompt. Using a RTX 3090 with Ubuntu 24, any thoughts? Could it be the new 4 Step LORA?

Aitrepreneur

This looks like a connection issue maybe a VPN or something else interfering

Aitrepreneur

Use the combo workflow of course, to get the best results

John Smith

nope not a VPN it does it on all three downloads pretty much in the same spot.

Me Myself and I

Thanks! Have you tried it with this workflow, though? I am testing it on "TWO TIMES" option and I can use one of those loras at a time with the first lora loader...

po vaj

QWEN+QWEN IS CRAZY GOOD! game changing!

Jibzorz

I still get poor images with the lora in 4 steps, but decent images with 12 steps and above. I'm not sure if the rgthree lora loader is loading it right (everything's updated). any thoughts?

theworldofanatnom

same issue, theres something up with the 4 step lora so i just switched back to 8 step lora

Rob113

I have installed with bat and update CUI but getting error.. Unexpected text model architecture type in GGUF file: 'qwen2vl'

Bobo.Stone

ive never used comfy before so i dont know yet how to trubbleshoot but i get this error after using the 1 click installer anyone know whats the issue? Prompt outputs failed validation: UnetLoaderGGUF: - Value not in list: unet_name: 'Qwen_Image_Distill-Q4_K_S.gguf' not in ['Qwen_Image_Distill-Q5_K_S.gguf', 'Wan2.2-T2V-A14B-LowNoise-Q5_K_S.gguf']

Jibzorz

go into comfyui manager, node manager and manually update or change ComfyUI-GGUF node to nightly version. worked for me.

Jibzorz

means the Q4 K S gguf file is not in the \model\diffusion_models. the Q4 version is preselected in the workflow, and if you downloaded the Q5 version, you have to manually select it in the unetloadergguf node.

Bobo.Stone

oh thank you so much now it worked

Jibzorz

no worries. if its you're first time using comfyui, i'd play around with much simpler workflows just to get the hang of the basics. though, this one is pretty awesome.

Rob113

Thank you sir. That fixed it for me as well. :-)

Me Myself and I

Looks like realism is improved by reducing the strength of the lora on first run (under "two times"). Tried turning it off entirely, but that wasn't ideal either.

Angelo Vildasol

I'm getting this same exact error message too

Angelo Vildasol

Pretty sure my version is all up to date

Angelo Vildasol

"qwen_image" doesn't appear to be in my type dropdown list

Angelo Vildasol

So it appears that the QWEN-ULTRA-MODELS-NODES_INSTALL.bat doesn't update the nodes quite right (in our case). So I just did a fresh install and that seemed to have worked.

Arthur

same here with the 4 steps

ak pak

close everything and run "update_comfyui.bat" in "ComfyUI_windows_portable\update\"

Angelo Vildasol

This is happening to me as well. The image comes out completely blurry and muddy with 4 steps. The 8 step lora looks MUCH better, but it still doesn't look anything like Aitrepreneur's results...

Angelo Vildasol

I'm not getting AiT's crisp stylized render, but I AM getting a damn good photorealistic result with the 8 step lora with not too much render time. Either way, color me impressed!

Aitrepreneur

yes but I can't provide support for it since I don't use apple products

Aitrepreneur

check if you get errors, maybe the lora is not loading correctly, hence why you need to use the comfyui nightly version as explained

Aitrepreneur

gguf models are located in the models/unet folder

Boris

This worked for me, thanks a lot

King Debs

cannot reshape array of size 7802545 into shape (18432,1728) Someone please help.

Aitrepreneur

just update comfyui with the update comfyui bat file inside the update folder and you change to the nightly build inside the manager like showed in the video

Angelo Vildasol

Do you happen to know if there are any other lora formats that might be compatible with this? I've only managed to find about a handful of QWEN loras so far.

Aitrepreneur

it's a fairly new model, be patient or train your own, I'm training a few right now, it's pretty good

uru team

help CLIPLoaderGGUF GGUF magic invalid

Ryan

I updated comfy and am using the nightly version and getting this cannot reshape array of size 24259885 into shape (3072,8448)

Teresa Greasley

I am getting the same error and updated comfyui, tried different Ksampler settings nothing is working

James

Yeah; I ran the nightly and it rolled something back and removed the qwen option from the cliploader. fortunately just going back to stable and it got restored.

Zikai

1) What file type or LoRa's can be used for WAN and Qwen? 2) Do you just keep the seed if you want to retain the same subject? for consistent subject generation for different things?

Gargamel

I've come across this a few times with 1 click installers. I use AVG security and have to temporarily disable it in order to complete the download. I'm not sure if its AVG specific or an antivirus issue in general, but that's what I had to do.

Killy_Blame

Seems promising for inpainting to or are the results bad?

Tetsuo Oshima

Hi K ! I can't remember if I said that before, in RGThree Settings at the bottom there is an option Groups > SHOW FAST TOGGLES IN GROUP HEADERS, so the Fast Groups Bypasser became obsolete, unless it applies to several groups

Timmeh

I can't see the link to download the updated workflow? I managed to download the original workflow a couple of days ago but not seeing it now.

King Debs

Hey, did you get this resolved, I'm still stuck

Bob

i2v is kind looping at the end, but my loop_count is set to 0, how do i fix it, i want the model stay in position after movement, for example i say "move to left" then she moves to left but at last second she walks back to right

Timmeh

Worked it out. For anyone else with this issue if you view the post from the Patreon home page the links don't exist but they appear fine when you browse the posts from the Aitrepreneur page directly.

Brudolf Wikswolle

I have a problem I cant run the workflow. Every time I try to run an error stating: Clip loader gguf unexpected text model architecture type in GGUF file "quen2vl". Does anyone knwo a solution for that?

Gardan

Thank you so much, I got it working after getting only blurry images (like I got also with the WAN 2.2 Image2Video). I finally figured out that I was missing some Loras and now it works. Thanks so much for your work, you're awesome. Also actually this was probably your funniest video, with the comments about WAN 2.2. ;)

HealingPaint

Try to close the program.... and run the comfy update bat file inside of your comfy folder's \update folder. That was the only thing that could fix the GGUF loader errors I was getting.

HealingPaint

Would it be possible to add one more module to this workflow to complete it; turn your final image into a video, like runi it through wan 2.2? That would be awesome.

Morpheus

Indeed, even with Q4 Loras QWEN apparently needs at least 10 step and cfg at least to 2.5. It gets decent results that way but still quite a soft image... a far cry from what WAN 2.2 can do quality wise. The only thing that is better with QWEN is prompt adherence. It's very good on that point.

Morpheus

When it comes to generate portraits of human faces with perfect realism, only ONE AI model can achieve that kind of perfection, some specially tuned Flux models do (not the base Flux model though, its quite "plasticky"). However... Wan can do all the rest Flux can't do well: like great general anatomy, very rich environments, powerful creativity etc.... Owen's quality output is not as sharp as Flux and Wan, also it can't render eyes that well, but the prompt adherence is excellent.

Orc99

Hi. I'm having this same issue. Its not quite a loop, but it looks like its trying to loop. Any solution to this?

Gardan

Check his previous youtube video for WAN 2.2, there you have the workflow for image to video, and you can copy that part over in comfyui (ctrl+drag to select that part and ctrl+c, ctrl+v in the new workflow)

Brudolf Wikswolle

thanks a lot for the advice, but I did that already and it didnt work out. I cant find the solution :(

Bob

i don't think its solution, but it works! in prompt, write frames for each move you want, for example: frame 0-50 drink juice. frame 51-81 put cut on table.

Mikael Svenson

Is there a way to share height/width between the two parts of the combo? To save time updating both t2v and i2v when changing aspects?

Aitrepreneur

The stable version got updated, just update it to the latest version now

Aitrepreneur

1) Each lora needs to be trained for each model, only lora trained on the wan model work on the wan model, same with Qwen or any other models 2) Possibly but the best way is just lora training or using something like the Kontext model

Aitrepreneur

Never had that issue, maybe use more detailed prompts

Aitrepreneur

The solution is indeed to update comfyui correctly using the update comfyui bat file inside the update folder, then re drag and dropping the workflow inside comfyui and making sure that qwen is selected in the clip loader node

Aitrepreneur

yes you can create 2 int nodes and then connect them to the width and heigh part of the nodes

HealingPaint

Ahhh I didn't know you could copy stuff from one workflow to another one. I'll try that, thanks.

Bob

i think i found best seed: 999999999999999 1039338885633958

King Debs

Hi, where in this workflow would i add the Juggernaut XL from CivitAi ? or generally, how do i modify this workflow for other LoRAs and Models from CivitAi ????????

Josh Gmit

different models need to be handled differently. You may get a workflow for SDXL or other Model-Architectures on CivitAi too.

King Debs

hi, thanks for the response. i read that there's different handling, but i dont think i have an answer. I have downloaded Juggernaut XL, its Base Model SDXL 1.0. on CivitAi it says it's a checkpoint. how do i use this in K's workflow???

Hank

I have used the RunPod installer and the Workflow for QWEN on a RTX 5090 and i get this: CUDA error: no kernel image is available for execution on the device 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. it does work on a RTX 4090

BeLiaL522280

So no matter what i do i keep getting this message "When loading the graph, the following node types were not found Upscale by Factor with Model (WLSH)" You mind passing discord inv and helpinh me out?The invitation under ur video doesn't work

Aitrepreneur

you can't use blackwell gpu with my template right now unless you are willing to do some manual install commands

BeLiaL522280

Some nick or how can i find you?

Zoran Kanti-Paul

I tried the Qwen image to image, every time I do it either comfyui just freezes up or my pc actually crashes.

FILM426

I'm getting this error, can anyone advise? mat1 and mat2 shapes cannot be multiplied (5292x1280 and 3840x1280) thanks guys

SpaceMonkey

I did the installation in runpod but when i try any workflow, it get stuck. It never finish. i did try the qwen text to image, qwen combo, ect. I hit run, is reach the ksampler and stay there...

SpaceMonkey

did you find a solution? i have the same issue

Mikael Svenson

Works like a charm. I also pulled in latest nunchaku and WOW, boy oh boy oh boy, what a speed improvement on my 4060. But something was off on actual generation causing bland images. Tested in the double floe. Question, why have two model nodes in the two step flow instead of connecting one model node to both samplers?

Aitrepreneur

Update comfyui by using the update comfyui bat file inside the update folder and update all the nodes by using the manager by clicking on update all, then you can restart and re drag and drop the workflow in again

Josh inmn

I recall someone asking for an auto prompter but I cant recall where someone in here though. I created an wan2.2 prompt generator to properly generate the correct style of prompts for the most realistic, and advance. It uses either LMStudio or OLLAMA for the back end. I have an exe and the source code on my github. Hope it helps someone. https://github.com/Wontfallo/Wan2-2_Prompt_Generator.git

AILab: MuseEngine

i managed to make all the issues and alert work by using chatgpt. DO the same if you have issues

John Jan

hi bro, iam getting blurry images as well. could you help me with fixing this

TrapHouseMag

How do you fix the white screen?

Sabsabinx

Same here. The first of image of the combos is very blurred, especially in the background. For example the text is totally unreadable on the signs for the default cyberpunk girl prompt.

Romercial

Same here, i can't figure out why. I haven't gotten consistent output quality since the WAN 2.1. It's always blurry. Or some sort of ghosting effect or sometimes the faces are pretty clean but then the background or the text on the foreground or on the shirts gets blurry.. Help us out, homie, please.

Airdropkazan

mat1 and mat2 shapes cannot be multiplied (784x1280 and 3840x1280) i tried several clip models but it didnt work for me. google says maybe its from the node itself.