KREA-2 GGUF Weights
Quantized GGUF weights for the Krea-2 Base model, optimized for low-VRAM and consumer hardware local workflows via ComfyUI.
GGUF MODEL FILES:
https://huggingface.co/realrebelai/KREA-2_GGUFs/tree/main
馃殌 Overview
Krea-2 is a Diffusion Transformer (DiT). Standard GGUF custom nodes will throw an Unexpected architecture type error natively. To run these weights, you must use our patched custom node fork.
馃摜 Downloading and Installation
1. Custom Node Requirements (Mandatory)
Standard loaders do not yet recognize krea2 architecture tags in GGUF metadata. Clone the following custom node into your ComfyUI/custom_nodes/ directory:
Repository: RealRebelAI/ComfyUI-GGUF_KREA-2
2. Required Support Models
Place these base model files in your standard ComfyUI models/ directory:
Text Encoder (CLIP): Qwen3-VL-4B-FP8-Scaled.safetensors (Load via
CLIPLoader, ensuring Type is set tokrea2)VAE: qwen_image_vae.safetensors (Load via
VAELoader)
3. GGUF Weight Installation
Download your preferred quantization file (e.g.,
Krea-2-Base-Q4_K_S.gguf) from the Files and versions tab of this repository.Place the downloaded
.gguffile into yourComfyUI/models/unet/(or equivalent GGUF directory) folder.Load the model in ComfyUI using the
UnetLoaderGGUFnode.
鈿欙笍 Workflow Notes
Ensure your conditioning setup is mapped appropriately to handle the stacked hidden state requirements of multimodal DiTs to avoid latent sequence mismatch errors.
Description
base ggufs for turbo and base
FAQ
Comments (25)
I installed everything you listed, RealRebelAI/ComfyUI-GGUF_KREA-2 etc. but the gguf cliploader doesn't have krea2 to choose from (the node even shows it's GGUF_KREA-2 node), am I doing something wrong?
EDIT: Turns out I just had to update comfyui itself, works now, thanks!
Well awesome! Great problem solving, i was going to say yes you just need to update comfy 馃憦馃憦馃憦
@realrebelai聽actually, the workflow ran, but the resulting image was nonsense, just colors and shapes like circuit board, I tried with gguf model Krea-2-Turbo-Q6_K.gguf. When I swapped back to fp8 non gguf model, the workflow works normally and produces proper image, so something to do with the gguf itself :S
@Bliney聽ill have to take a look, they all ran fine in my testing im not sure what happened, did you use the wrong workflow with the wrong model? Sounds like that couldve been it. If you drop your outout it gave you back into comfy itll show you the exact workflow settings than ran on that image. That will confirm if you used the wrong settings then we can go from there
@realrebelai聽I found the culprit, I had the clip from CLIPLoader (GGUF) from the GGUF_KREA-2 custom node, but when I swapped back to normal CLIPLoader it gives a proper output
@Bliney聽nice! i had retested it on my break an got good results so i had came back to let you know but it seems like youve fixed it! so thats great
can you make it run with a gguf qwen 3 vl 4b model ?
Hi there, I think @realrebelai will add support too, until then you can use my fork of the ComfyUI-GGUF nodes (described here) https://civitai.com/articles/31860/run-krea2-as-a-gguf-in-comfyui
You can keep Realrebelai's workflow and just use the load clip gguf node, it will be compatible.
@molbal聽how were you able to introduce a gguf into the workflow?!?! thats awesome man!
@molbal @realrebelai sorry guys i am noob to comfyui, i wish if you could update the workflow and add loras as well
@cabsluke2022942聽loras are simple brother, if you cant figure that out your cooked lol you need to just string the model into the left of the lora node then string the right side of the lora node into the ksampler. thats it :)
@realrebelai聽Most of we work was done by city96, the original maintainer of the comfyui-gguf nodes
@molbal聽i know that, im saying the gguf encoder seem to be missing key layers to encode these newer models properly. Every test i ran failed with mismatch shapes
@realrebelai聽oh that! Those layers are apparently not used during CLIP so I just pretend those layers are there, this is the implementation: https://github.com/molbal/ComfyUI-GGUF/commit/cd2c9bdca2ee0f83760abd2822d9525b30c4fe6d#diff-86393d83b45516ef8888170f0149c4c74625a15d65eb084d7f74df2bbf41d1f1
@molbal聽interesting workflow, i got it working, but i thought it would be faster when using gguf text encoder
@cabsluke2022942聽I don't think its faster, it might help for low VRAM though, because GGUF files contain built-in metadata that allow ComfyUI to easily identify and split specific model layers between the GPU and CPU. However, safetensors are just raw data, and ComfyUI must reconstruct the model structure and manage memory allocation during the load process if it wants to offload parts of it to system memory from VRAM.
Hey - great work with the GGUF versions - I have managed F2K4b and F2K9b versions with your workflows so far. Works well.
Trying the Krea2 GGUF and put the custom node in via git (shows up in directory) and comfyui is latest version and restarted. Trying the lower Krea-2-Base-Q3_K_M.gguf - it shows up in your workflow. When running I get this error immediately;
"Unet Loader (GGUF)
ValueError: Unexpected architecture type in GGUF file: 'krea2' "
I have gone through every step several times and can't get past this error. Can't think of anything more to try - hope you can help.
Again thanks for the work to open up all these models to lesser (GPU VRAM) mortals.
Hi there, same problem here, did also what you described but no luck, having this error no matter what I do, clearly I'm missing something there, maybe anyone have some idea what it may be ? Thank you!
did you install my nodes?
I have installed GGUF from comfy manager and your GGUF_KREA-2 from git. After opening workflow and running it I have this error, I can see also the nodes in custom nodes menu, but even if I pick from there the Unet loader that says GGUF_KREA-2 (bootleg) it gives me the same error while processing the prompt (clip text encode). Since some of the things are very new to me regarding Comfy I may be doing some basic stupid mistake here. Thank you!
@a90201740聽no no ! Its never stupid brother! This is a hard learning process dont beat yourself up! I just hate when i cant help someone 馃ぃ馃挃
Maybe you have conflicting dependencies from other custom nodes? Sometimes that can interfere
I will check it again with fresh copy and only those nodes and see if this helps. There is a possibility this may be the case because I have many. Thank you!
i was asking google ai mode about this new mrflow trick to speed up image generation, it mentioned your name and said you did use it, is that true ?
i did! thats insane! i havent posted it to civitai but i will shortly as im about to upload another model i just quanted called sefi-image. give me about 20 minutes maybe 30 and ill have it up for you brother!
I have been experimenting with Turbo Q3 K_M, with great success. Thank you!