I'm using Forge UI and getting a similar error (VAE state dict gives weight dimensions as 3072x384, but the number of in channels when the model is initialized is 64).

Updated and same Error.

I have encountered the same problem.

After the update, it can now run normally. Thank you for sharing

For anyone looking for this: YarvixPA/FLUX.1-Fill-dev-gguf · Hugging Face 👍😁👍

Is GGUF better then just the FP8?

@runebloodstone Yes, generations with GGUF 8 are almost identical to FP16, whereas FP8 is less accurate.

tested this against SporkySporkness/FLUX.1-Fill-dev-GGUF quants and the differences are very subtle with the same seed. I would guess any flaws are in the original model (or maybe the frontend you're using).

stack a nsfw lora

same here, I already update to the newest comfy, but I still get pixelated image, don't know why

I have the same thing. The image gets very muddy, consisting of dirty pixel mush. I used the standard fill workflow.

Can confirm that the problem is also present on the GGUF models (2 of the made gguf models have this).

@Dainzh That's right, I was just using gguf. I took Q4 and Q5 from here: https://huggingface.co/SporkySporkness/FLUX.1-Fill-dev-GGUF and they both have this problem. Is there a normal gguf without this problem?

Seems like not a gguf problem, or fp8 problem or any other problem model related, because I've just tested with regular fill dev mode which is 22 gb and it has the same problem. Something tells me this is due to the ComfyUI problem itself, and I don't really know what might be causing it.

@superuser111 Are you using any custom nodes by chance?

@Dainzh Very sad, hope the problem is resolved soon.

@Dainzh I used the standard workflow taken from here https://civitai.com/models/970162?modelVersionId=1088649, only replaced flux model node with gguf

anyone solved the noisy problem?

@superdumbkitten What GPU do you have?

Thanks OP it worked! For those who are struggling try using the image provided by OP for testing.
Also found this guide below which gives details on expected parameter values.
https://comfyui-wiki.com/en/tutorial/advanced/flux-fill-workflow-step-by-step-guide

@Dainzh 8gb 2080s

For thosee who still struggling for the noisy image, adjust the cfg to 7or8, the cfg is 1 in detault workflow.

SDXL outpainting fast and quick less nodes https://youtu.be/YNn-MUCRs60?si=yVtQR4eE-nRPC9xq

Just use SDXL https://youtu.be/YNn-MUCRs60?si=yVtQR4eE-nRPC9xq

@1girls No thanks, it's like using a push-button phone instead of a smartphone. Yes, it works too, but it has been outdated.

https://www.youtube.com/watch?v=wvAFmYF7qgA

Workflow in the description. Includes outpainting and inpainting.

@13_february wrong concept.. flux is overhyped. technically you dont need it if you are good with the tools in hand

@junkyard001 loads of workflow nodes from third party repo is trash.. more nodes doesnt mean you are good at it.. i dont like youtubers workflows i make my own

@1girls Even so, I didn't ask for an outpaint workflow based on 1.5 model🤷 It cannot work out small details efficiently. Yes, it can expand an image with a simple background, but highly detailed images are not it's strong suit. Flux didn't get its overhype for nothing. Its main feature is high consistency in details. With it, you don't get a hodgepodge in details like in the 1.5 and SDXL models when you you work on high-resolution art illustrations.

hi, what do you mean unet folder? isn't this suppose to go in /models/diffusion_models/ ??

use gguf https://youtu.be/mDTKQIClUMs?si=rGIC_kIU_W2xVvoJ

Load Diffusion Model

you can try custom workflow from sebastian kamph ACE++ Character Consistency without training - Only 1 input image | Patreon

Getting the same issue. Can't use ComfyUI as it shows the below error. I wonder if any tried this with WebForge NEO.

CLIPTextEncode CUDA error: no kernel image is available for execution on the device Search for `cudaErrorNoKernelImageForDevice'