Home
Creators
Search Recent Missing 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

FLUX2 KLEIN THE MOST VERSATILE FREE NANO BANANA COMPETITOR RIGHT NOW! (Patreon)

Published:
2026-02-22 17:59:14
Edited:
2026-09-03 00:54:05
Imported:
Tags:

Downloads

Content

EDIT 3 September:NEW RUNPOD CUDA 13 TEMPLATE!

I’ve spend the last few days, completely rebuilding my RunPod ComfyUI template from scratch.... The previous template was getting very outdated and was starting to cause problems with newer GPUs, ComfyUI versions and recent workflows, so please use this new template from now on.

The new template uses CUDA 13, supports newer GPUs like the RTX 50 series, comes with SageAttention preinstalled, automatically keeps ComfyUI updated and works with my new RunPod installers.

HOW TO USE THE NEW TEMPLATE

I made a quick video showing exactly how to deploy it and use the installers:

👉 https://youtu.be/9wlUXXyy9mE

👉 RUNPOD TEMPLATE LINK

IMPORTANT: When choosing your GPU, use the CUDA 13.x filter on RunPod

The old RunPod template is now completely retired, so use this one for all new installs going forward.To use the installer use this command:bash "KLEIN_EDIT_ULTRA-AUTO_INSTALL-RUNPOD-NEW.sh"I WILL SPEND THE NEXT FEW DAYS UPDATING ALL PREVIOUS RUNPOD INSTALLERS TO THE NEW TEMPLATE SO IF THERE IS AN INSTALLER THAT I FORGOT LET ME KNOW.Hey everyone! I've created a 1-click installer for the new video: THIS IS THE MOST VERSATILE FREE NANO BANANA COMPETITOR RIGHT NOW! called “FLUX2 KLEIN". An insanely good model that can do everything: Text to image, image editing, inpainting, outpainting and all of that with super fast generation speed on low vram! AND it keeps getting stronger AND stronger because the LoRA ecosystem exploded, and with the right workflow you can do EVERYTHING! JUST INSANE! 

You can check out the video right here: https://youtu.be/N7j3lvenDcU

The workflow: https://www.patreon.com/posts/151396151

The installer of course automates the entire install process for comfyUI and the models downloads!

LOCAL INSTALL!!! RECOMMENDED TO START FROM SCRATCH

Download the KLEIN_EDIT_ULTRA-COMFYUI-MANAGER_AUTO_INSTALL.bat

Run the bat file

Wait for the install to finish

The webui will launch, load the KLEIN_EDIT_ULTRA_WORKFLOW.json file and there you go😎

DON'T FORGET TO GO UPDATE COMFYUI BY GOING INSIDE THE UPDATE FOLDER AND RUNNING THE COMFYUI UPDATE BAT FILE!!!!

IF YOU ALREADY HAVE COMFYUI INSTALLED USING A PREVIOUS INSTALLER:

Download the KLEIN_EDIT_ULTRA-MODELS-NODES_INSTALL.bat and place it in the ComfyUI_windows_portable\ComfyUI folder

Run the bat file

Wait for the install to finish

???

Profit 😎

IF YOU GET AN ERROR WHEN LAUNCHING COMFYUI AFTER USING THIS INSTALLER SAYING "AssertionError: Torch not compiled with CUDA enabled" it's ok don't panic.

Just go inside the python embedded folder, click on the folder path, type cmd press enter, this will bring the command prompt window and there type:

python.exe -m pip uninstall torch

python.exe -s -m pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

--->replace cu128 by your cuda version (if you don't know your cuda version just open a new cmd window and type nvcc --version it will then tell you at the end Cuda compilation tools, release 12.8, V12.8.93 or something different like 12.6, in that case just use cu126 instead of cu128)

IF YOU ARE USING RUNPOD:

Create an account if you haven't already: Runpod 

Click on Pod (on the left side) then click deploy

When selecting your GPU on RunPod, make sure you click on Additional Filters → CUDA Versions → CUDA 13.0Then choose a GPU, click here for the comfyui template (aitrepreneur/comfyui:3.0.1-cuda13)

Then click deploy, wait for everything to finish and then click on port 8888 and then drag and drop the KLEIN_EDIT_ULTRA-AUTO_INSTALL-RUNPOD-NEW.sh file on the left side of the UI then click on the "Terminal" icon on the right side on the UI

Copy and paste this line then press enter:

bash "KLEIN_EDIT_ULTRA-AUTO_INSTALL-RUNPOD-NEW.sh"

Wait for everything to be installed

Then go to the port 8188

Load the KLEIN_EDIT_ULTRA_WORKFLOW.json workflow and there you go😎

Also use the command:

tail -f comfyui.log

inside the workspace/logs to get the real time logs

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

ED-3D

Thank you

Paul Mazurkiewicz

Thanks as always for the best workflows in the industry!!

Ádám Simon

Please, can someone help me? I really don’t understand what I’m doing wrong, even though I’ve been trying for days. I want to create a LoRA of a woman, but it’s extremely difficult to consistently generate the exact same model (fixed seed, fixed base prompt). With Z-image Turbo it’s basically impossible, so I’m using the regular Z-image with the workflow below. Am I doing something wrong, or should I be using a completely different base model instead of Z-image? (I’m about to try Klein Edit next.) Thank you!!

Aitrepreneur

Z-image has unfortunately fallen short when it comes to training...at least as of right now, everyone is extremely disappointed with its lack of correct training (me included) so much so that a lot of people are moving toward Klein since the model is super ultra versatile and super easy to train, I'll make the next video on that, since I really liked the model. I still really like Z-image, I love that model but when it comes to training, it seems like the community has kinda abandoned working on it until further notice or new breakthrough...

H-BOX

Thank you sir I love all your work. Stay healthy and have fun always !

Anthony Harris

When I installed the missing nodes onto an current install of ComfyUI, I get some missing customs nodes which I can't find in the nodes manager. Including FastFilmGrain, FastLaplacianSharpen and LayerUtility: ImageRemoveAlpha. Also I updated ComfyUI to the latest version

Tark Morris

hey thanks for this, its amazing as usual however... to create an image its taking me 75 seconds, definitely not the 2-4 seconds you suggested, I have a 5070 ti asus so definitely not a PC issue, I ran the auto.bat file to just install the models etc... everything works fine but its just taking way too long to generate an image

Whistler

1. If you are using too big models then it have to offload stuff which makes things take much longer, so see if you can find something smaller if that is the case, the logs when running should tell you. 2. If not, or even if remove the clean vram and clean cache nodes that are hidden under the subgraphs. Otherwise you need to copy the models to vram before every generation! I do not know why he always put those in makes it much slower. But I guess it makes it a safer to not get OOM errors, with loras and stuff if you have very tight with vram. But ComfyUI have manual buttons for it if you need to clear it at some point. 3. Look into what Attention Method you are using, if you do not know you probably are using some old one that is really slow. This can cause oom errors also if you are using a bad one. (I have not run the bat files though so do not know how that set things up)

Whistler

Don't know why you can't find them but first one is vrgamedev and the second one is ComfyUI_LayerStyle search them in the manager and install them. You can also git clone them in the custom_nodes folder but try manager first it is easier.

Olivia C

Did anyone find qwen3_8b_fp8mixed_abliterated.safetensors ?

Tark Morris

yeah nah I was running a .bat for the ltx-16gb instead of the normal nvidia.bat file lol, now it generates in less than 7 seconds, all good but thanks for your input!

Can Turan

my generations take up to 70 seconds just like the others. I'm using false for gguf and didint change anything else. I have 5080 and using nvidia bat

Anthony Harris

Thanks for the reply, I decided to go with a fresh install of the Flux2 Klein and then install the Z-Image Base, Qwen and Z-Image Turbo into that install and delete my others. Actually it worked out good as I cleaned up my AI stuff and freed up space too

Ramjet

This is working well apart from a lot of the images look "dirty". if i make an image of a person, it looks like they have just come from under a car and have grime on their skin. Its a little better if i turn off most of the Lora's. GPU 5090.

Alpha Ghost 47

Hi, firstly thanks for all your hard work that clearly goes unappreciated as per comments. Can you inpaint using control net with faces. I make a lot of youtube thumbnails etc and been looking for something that does face inpainting since controlnet with sd 1.5. Thanks in advance for answer and keep up the hard work!

Sean Rowe

This would be awesome! Facefusion is a decent option if you get something close but what to improve it.

Sean Rowe

For some reason on the Klein edit workflow keeps producing images that are 1024x1024 despite the original image size. Isn't the image supposed to retain the landscape or portrait style when generated? Some of the generations are coming out distorted because it's widening the image a lot and making thing look strange.

Whistler

Set "Enable MANUAL IMAGE SIZE" to no. Then you can choose if you want "Enable AUTO RESIZE" on or not to pick Megapixels or keep the original size.

Keen

Hi! New subscriber. I downloaded the installer to a second drive on my computer because it has much more free space than my main drive. I used the install bat, but I have gotten a ton of warnings during the installation along the lines of "The script websockets.exe is installed in 'P:_______' which is not PATH. Consider adding this directory to PATH..." Any advice? Do I need to anything? If yes, can you talk me through what to do? I have a copy of all the warnings.

Keen

Wondering if that is why I got this error on my first attempt? Prompt execution failed Prompt outputs failed validation: UnetLoaderGGUF: - Value not in list: unet_name: 'flux-2-klein-9b-Q8_0.gguf' not in ['flux-2-klein-9b-Q4_K_S.gguf'] # ComfyUI Error Report ## Error Details - **Node ID:** N/A - **Node Type:** N/A - **Exception Type:** Prompt execution failed - **Exception Message:** Prompt outputs failed validation: UnetLoaderGGUF: - Value not in list: unet_name: 'flux-2-klein-9b-Q8_0.gguf' not in ['flux-2-klein-9b-Q4_K_S.gguf'] ## Stack Trace ``` Error: Prompt execution failed at ComfyApi.queuePrompt (http://127.0.0.1:8188/assets/api-Cllpqoko.js:4:6373) at async ge.queuePrompt (http://127.0.0.1:8188/extensions/ComfyUI-Easy-Use/assets/extensions-DeJF1X-A.js:1:158311) at async PromptService.api.queuePrompt (http://127.0.0.1:8188/rgthree/common/prompt_service.js:88:28) at async ComfyApp.queuePrompt (http://127.0.0.1:8188/assets/dialogService-45wv8NTN.js:177:17013) at async app.queuePrompt (http://127.0.0.1:8188/extensions/rgthree-comfy/rgthree.js:478:24) at async Proxy.function (http://127.0.0.1:8188/assets/GraphView-9xTw3uVt.js:8:78471) at async http://127.0.0.1:8188/assets/useErrorHandling-r0vTEz20.js:1:487 at async Proxy.execute (http://127.0.0.1:8188/assets/dialogService-45wv8NTN.js:2:89608) at async queuePrompt (http://127.0.0.1:8188/assets/ComfyQueueButton-Qkiy8R1k.js:1:2840) ``` ## System Information - **ComfyUI Version:** 0.14.1 - **Arguments:** ComfyUI\main.py --windows-standalone-build

Toneski BG

pose doesnt seem to work, ive changed the prompt 2 many times and same result. The Character retains original pose

Whistler

Yes it is hard to get it to work. Try and give a short explanation of your subjects is what seem to help the most for me, even say if you are using control net otherwise describe two people. So for example "Change the pose of the woman dressed in a white top and denim pants from image 1 to the pose of the control net in image 2."

Whistler

Click the lists under Models Selection at the top and pick your models. If you do not see them they are not in the correct folder. The first warning is about Environment Variables in Windows. I'm unsure if you need to set them or not, I have not set them but I'm using a portable version.

Keen

Thanks, Whistler! I downloaded the missing models from https://huggingface.co/Aitrepreneur/FLX/tree/main and added them to my unet subfolder. And added a prompt to Klein Text to Image Combo... And it is running! And in the time it took me to write that, it was nearly done, and it just finished now. Thank you so much for helping me get this far!

Replicant

Loading aborted due to error reloading workflow data Error: invalid connection Can't even load workflow, any help? Thx

Aitrepreneur

which gguf though? you can try fp8 as well with 16gb of vram

Aitrepreneur

I haven't tried but i can try on my side and see if that works, not hard to modify the workflow yourself though, you got the edit example where you got the controlnet node, just replace that with depth or something like that

Aitrepreneur

Exactly like Whistler said, controlnet doesn't work as well with klein as with Qwen image edit but it still works if you take the time to detail the pose just like I showed in my video

Keen

So I got to LTX-2 working with Whistler's help... and now it keeps giving me an ffmpeg error. Have no idea how to get out of that hole -- even tried installing a node FFMPEG "A custom node for ComfyUI to save image sequences as video files using FFmpeg. Supports various codecs, audio muxing, and in-node previews." ID 2108 and title "Save Images to Video (FFmpeg)," but still getting the same error. Help, please! More details in the LTX-2 thread. ----- Update -- wish me luck! I think I got FFmpeg installed!

Kaiser Lowendachs

Clip loader error - I seem to get this with most of your flux workflows model starts then displays a purple box around the Load Clip box. with this text report. I downloaded from you hugging face the exact clip and dropping in the comfyui/models/clip folder - no change. Any solution gratefully recieved Error(s) in loading state_dict for Llama2: size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([151936, 4096]) from checkpoint, the shape in current model is torch.Size([128256, 4096]). size mismatch for model.layers.0.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.0.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.0.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.1.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.1.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.1.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.2.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.2.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.2.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.3.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.3.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.3.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.4.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.4.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.4.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.5.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.5.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.5.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.6.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.6.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.6.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.7.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.7.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.7.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.8.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.8.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.8.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.9.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.9.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.9.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.10.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.10.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.10.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.11.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.11.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.11.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.12.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.12.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.12.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.13.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.13.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.13.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.14.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.14.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.14.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.15.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.15.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.15.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.16.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.16.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.16.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.17.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.17.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.17.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.18.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.18.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.18.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.19.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.19.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.19.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.20.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.20.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.20.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.21.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.21.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.21.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.22.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.22.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.22.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.23.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.23.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.23.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.24.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.24.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.24.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.25.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.25.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.25.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.26.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.26.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.26.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.27.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.27.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.27.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.28.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.28.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.28.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.29.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.29.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.29.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.30.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.30.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.30.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]). size mismatch for model.layers.31.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.31.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.31.mlp.down_proj.weight: copying a param with shape torch.Size([4096, 12288]) from checkpoint, the shape in current model is torch.Size([4096, 14336]).

Brian

I am getting this error when running. Anyone know what is happening here? mat1 and mat2 shapes cannot be multiplied (677x12288 and 7680x3072)

Whistler

When I get that error it is because the text encoder model that is not compatible with the Image Model.

The Irish Canuck

This is a wrong CLIP model version error. The workflow expects Llama 3 (vocab size 128,256) but the file downloaded is Qwen2/other model (vocab size 151,936) with different MLP dimensions. The file placed in the clip folder is simply the wrong model — it's architecturally incompatible regardless of the filename. Fix: You need to download the correct CLIP/text encoder the workflow specifically requires. Check the workflow's readme or description for the exact model name and source. Common Flux text encoders are: t5xxl_fp16.safetensors or t5xxl_fp8_e4m3fn.safetensors (T5) clip_l.safetensors (CLIP-L) Long-ViT-L-14-FLUX.safetensors For LLM-based CLIP nodes: specifically Llama-3 variants (not Qwen, not Llama-2) T5-XXL (most common Flux requirement) https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main That repo contains t5xxl_fp16.safetensors, t5xxl_fp8_e4m3fn.safetensors, and clip_l.safetensors — the standard trio for most Flux workflows. For LLM-based CLIP nodes requiring Llama 3 (like the one in your error), you'd need a Llama 3 8B model, but those require a Meta/HuggingFace access request: https://huggingface.co/meta-llama/Meta-Llama-3-8B The key detail from your error is the model expects vocab size 128,256 which is definitively Llama 3, not Llama 2 or Qwen.

Richard Curtis

The last two *.bat releases you've made (thank you BTW!) are default installing in Windows/System32...i trust that's from the batch file. Would appreciate a default install location where I dropped the .bat. Thanks again!

a pretty parrot

what operation are you doing? on edit in 4090 runpod i am ranging from 7-15 secs depending on image size (i modified to keep original size).

a pretty parrot

yeah mine tend to come up a bit "dirtier" than zimage, but the speed is pretty bonkers and quality i really cant complain about

a pretty parrot

it would be easier to know what you are trying to do with the faces. if i remember right they do map facial features, so if youre doing something like a closeup face and want to change expressions it may work. i would also have the expression in the prompt to help it. and actually i think aitrepreneur did a video or post for sillytavern. the character expression part of that may be exactly what youre looking for. you can also just swap the models out of that (no idea what he used for image models) with whatever model you want. i made a dataset tool +oneshot lora workflow that could be good depending what youre doing

a pretty parrot

i use a oneshot lora generator workflow and i made a dataset tool to easily pick the best from each run. consistency lora helps. sometimes your base image may seem good, but it just does not handle changing the angles well

Michael Young

Last months didn't work for me, and now this doesn't either...bummed, these used to work so perfect from you. Prompt execution failed Prompt outputs failed validation: FastLaplacianSharpen: - Required input is missing: images It wont generate anything, so then all the sharpening and then the other stuff tosses errors.

Vyacheslav

Thanks for the great WF, as always! But could you please update it, so the inpainting mode works through the crop/stich node? So we can inpaint 4k images without loosing quality :) That would be awesome!

N. M.

Getting this too. Update: I used the batch file to install the dependencies (not the full install; I already had Comfy where I wanted it). It still didn't work. So, I grabbed a fresh download of the workflow, opened it, and it worked. I'm using the latest version of Comfy. 1. Install dependencies with KLEIN_EDIT_ULTRA-MODELS-NODES_INSTALL.bat 2. Download and open the workflow again. 3. ??? 4. Profit!

DLF

Thank you for this. I can save some hair :) BTW, what text encoders are woking for you?

Keen

Have been experimenting with both Flux2 and Klein. I have a question regarding Loras that are available on-line, and another about Loras I could make for myself. 1) Does the Lora I add for either Flux2 or Klein have to have been trained for use with those? I have seen a lot of sites offering Loras, but I have not figured how to tell which Lora could be added to either of the workflows we got in the bats. 2) There is a Lora trainer installed with the bat for text-gen UI. Is there is a tutorial here on making Loras, or additional information about how to make a Lora for use with a specific workflow, if Loras do have to be tailored for use?

Whistler

Qwen3 8b is what you want. Unsure if all of them work though. Currently I use qwen3_8b_abliterated_v2.safetensors that one is uncensored, not that I have ever notice the regular one was not working for something. But out of principle I do not want to use censored things.

Whistler

1. Yes it need to be made for the model you are using. Might be some exceptions but something will probably break or be worse anyway. If it is just a newer release of a model maybe it can work, ask Googles AI it will know if it does or not. 2. Aitrepreneur have made several videos about it. I have never done any though, pretty sure you use AI Toolkit for it and not ComfyUI. So he has probably some other bat files in those videos.

a pretty parrot

for lora, his other videos with aitoolkit should apply fine. i havent done in awhile, but i think big thing is make sure your input images resolutions match what you are training for. not sure what you mean about the text-gen ui, but aitoolkit im sure is already setup to handle flux2klein

a pretty parrot

what mode/subflow are you having the issues with? i can try to reproduce and tell you how to fix EDIT: i see where you are likely using them on the text2img subflows. the "images" are connected to the output from sampler subgraph, so what i would do first is: 1) click the FastLaplacianSharpen node, and press CTRL+b on it to bypass it (makes it purple) 2) try to run workflow, may not be pretty but get it outputting something 3) the FastFilmGrain might give same error, do same CTRL+b to that I believe those are from VRGameGirlDev so it could be an issue with that. When you start Comfy, look at the logs to see if there are any errors.

a pretty parrot

+1 to this, happens to me on occasion too. are you running local or runpod?

Robert Bersch

Ok I got this installed and got the workflow up but everything stops at 28%... I think I need to enable CUDA cause I get a warning "[Crystools WARNING] CUDA is available, but torch is using CPU." Anyone have an idea how I do that? I see that my GPU is always at 0%...

LW

Hi K, thanks for the new Klein workflow. Unfortunately, I am getting a message saying that I have not got the following custom nodes installed: FastFilmGrain and FastLaplacianSharpen I have updated comfyUI to the latest version but I still get the message and Comfy Manager doesn't know anything about them, apparently. So, how should I install these missing nodes?

LW

I opened ostris AI-toolkit but I didn't see an option for training loras for Klein. Is this something in the works? Or is a different lora trainer needed?

Robert Bersch

My CUDA version is "nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2025 NVIDIA Corporation Built on Wed_Aug_20_13:58:20_Pacific_Daylight_Time_2025 Cuda compilation tools, release 13.0, V13.0.88 Build cuda_13.0.r13.0/compiler.36424714_0" Should I use cu126? (trying that now)

Whistler

If it says you have 13.0 so you should use cu130

Paul Czaban

There should be an option. Update it. Unfortunately training times are way too long for me... I'm waiting for a guide

Ana Herrera

Will these auto setup work on Simplepod?

Robert Bersch

got it working after a few days.. missing some things like ffmpeg and something else. Also was using the wrong start bat. Now I need to find a good site or video about writing prompts for this cause the videos are pretty wonky.

So You Think You’re an Artist

hello some feedback: I was confused by your preview image which shows video motion, and your comaparison to nanobanana (which also does video), but after watching your youtube seems like this workflow does NOT do video. anyways, that threw me for a little loop. thanks for making the content you make. cheers

John Cannan

Flux2 Lora training? Just sayin’ ;)

ZenzeosArt

For image editing Klein 9B seems fine, but I dont get any better results from it than with ZIB or ZIT. Is it the new "king" because its multimodal? Or what are advantages here? ZIT gets more Styles and has superior LORAs, ZIB can get more creative stuff, ZIT better quality images,... Klein just for editing?

Anders Madsen

would be nice to get the option to choose where to install it, damn annoying get denied errors all the time cause its in system32 library

Alan Smile

All of his [...]_AUTO_INSTALL.bat files are designed to install exactly where they're run from.. so if you wanted to install it into d:\my_stuff 1. Open your D: drive and create a new folder named my_stuff 2. Drag and drop (or copy and paste) your (whatever-it-is)_AUTO_INSTALL.bat file into that new d:\my_stuff folder. 3. Run it: Double-click the .bat file from inside d:\my_stuff The script will detect its new home and download and install everything right there, inside d:\my_stuff\ComfyUI_windows_portable .

Antacker

I have comfyUI already installed but the KLEIN_EDIT_ULTRA-MODELS-NODES_INSTALL does absolutely nothing... I run it, press 1 (or 2 or 3 whatever) and nothing happens.

Jon Force

Hi All, quick question, I have this working fine and it's great, however I don't know how to change the base model on this workflow. Like there are some different checkpoint models I'd like to try using, but on the workflow, I can't tell where it's pulling the model from, there's just a node that says "get model" and you can't change anything there. How do I change the base model for something else within the workflow?

Wielobel

Not sure why, but evewrything i generate looks abysmal. Like hands coming out of legs etc. I followed the instruction from scratch, and it works, but the quality is like something from analog horror :| no idea what the heck

Wielobel

don't run it as administrator, that way it will install in the bat file location.

Gujz

Hi, I'd like to ask for some advice. When I try to install ComfyUI via RunPod, I first let ComfyUI update and then reset it. Every time I press F5 to refresh the page, I get an "Access Denied" error (403) or a "Access Denied" message. Has anyone else had this happen?

Aitrepreneur

Have you also tried the new 2.3 version with the new worfklow?

Aitrepreneur

Yes, I got that too, it appeared recently, the problem seems to be the browser and/or some extensions you got. If you take the url and open it in a private window or in another browser, it will work. Not quite sure what is causing this

Anthony

Sir, I am really impressed with this flux klein workflow, stunning image generation and photorealistic. Thank you very much ✨️.

XxWalla_WallaxX

If your having trouble and getting errors with FastFilmGrain and FastLaplacianSharpen I have found a fix. first go to the folder using the path: "ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-vrgamedevgirl" then open the requirements.txt file with notepad, delete all text inside and replace with: "kornia librosa imageio torchcodec" then save and close the file. Then go to the folder path: "ComfyUI_windows_portable\ComfyUI\custom_nodes" and open the windows command prompt in this folder by typing cmd in the file explorer address bar. Once open type "..\..\python_embeded\python.exe -m pip install -r comfyui-vrgamedevgirl\requirements.txt" and hit enter to run the command. Problem should be fixed after that.

XxWalla_WallaxX

If your having trouble and getting errors with FastFilmGrain and FastLaplacianSharpen I have found a fix. First go to the folder using the path: "ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-vrgamedevgirl" -then open the requirements.txt file with notepad -delete all text inside and replace with: "kornia librosa imageio torchcodec" -then save and close the file. -Then go to the folder path: "ComfyUI_windows_portable\ComfyUI\custom_nodes" and open the windows command prompt by typing cmd in the file explorer address bar. -Once open type "..\..\python_embeded\python.exe -m pip install -r comfyui-vrgamedevgirl\requirements.txt" and hit enter to run the command. Problem should be fixed after that.

Brudolf Wikswolle

Can you do the mask overlay with this model too? So that you can specifically target the objects and things you want to change in the picture?

Anders

im getting, Prompt execution failed Prompt outputs failed validation: FastLaplacianSharpen: - Required input is missing: images

Ivan Maya

im getting: • downloading detail_slider_klein_9b_20260123_065513.safetensors curl: (22) The requested URL returned error: 404

Kyle Williams

Please help: Installation is not finishing, showing error 401, therefore getting 4 errors "missing nods"

Gujz

Hi, I'm getting an error about missing nodes: FastFilmGrain, FastLaplacianSharpen. I'm using RunPod. Do you know what to do about this?

Aitrepreneur

those error are ok, you can just bypass the film grain and sharpen nodes

Miles

Hi everyone, I'm new! sorry but I'm, (*sigh*) running Comfy in a docker container on a linux box. the models nodes installers - can the RUNPOD one work for my situation safely? I'm asking where I should run it from - just the comfyUI resources /root dir? is there a manual method? I have installed all missing custom nodes and downloaded the models until those errors stopped. But I am still missing something obviously. please help Errors: Error(s) in loading state_dict for Llama2: size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([151936, 4096]) from checkpoint, the shape in current model is torch.Size([128256, 4096]). size mismatch for model.layers.0.mlp.gate_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]). size mismatch for model.layers.0.mlp.up_proj.weight: copying a param with shape torch.Size([12288, 4096]) from checkpoint, the shape in current model is torch.Size([14336, 4096]).

Untamed

I have the same problem, I can't seem to make it follow any controlnet pose.

Mark Burns

The JupyterLab on Runpod had this to say about the configuration today: ──────── Installing locked Torch stack ──────── Looking in indexes: https://download.pytorch.org/whl/cu121, https://pypi.org/simple ERROR: Cannot install torch==2.4.0+cu121 because these package versions have conflicting dependencies. The conflict is caused by: The user requested torch==2.4.0+cu121 The user requested (constraint) torch==2.10.0+cu130 Additionally, some packages in these conflicts have no matching distributions available for your environment: torch To fix this you could try to: 1. loosen the range of package versions you've specified 2. remove package versions to allow pip to attempt to solve the dependency conflict ERROR: ResolutionImpossible: for help visit https://pip.pypa.io/en/latest/topics/dependency-resolution/#dealing-with-dependency-conflicts root@ae930378e317:/workspace/runpod-slim/ComfyUI#