🐝 Please send Leinad882 all your buzz! 🐝
All credits go to Leinad882 on the Civitai discord for this. He spent 2 days non-stop working on this for me. I had to share it with the rest of you.
Why?
Many model merges, especially SDXL realistic model merges, are plagued with watermarks. Many watermarks can't even be negated. This is a workaround until we have better solutions or potentially an adetailer model. I created a bounty for this btw if anyone is interested: https://civarchive.com/bounties/1168
How to Use:
Specify if you're generating or loading an image: In the boolean switch on the left, choose 0 for loading an image, and 1 for generating.
0 - Load
1 - RenderUpload an image or start generating: Add an image or start configuring your loader and ksampler settings to start generating
Enjoy not having watermarks! (especially most SDXL plagued merges)
Utilizes the following custom nodes:
Enjoy!
Description
FAQ
Comments (11)
When you download this currently, it is just the image in the preview. Is that correct?
Yep, that is correct, the image contains the metadata. Just have to throw it in :)
I´m getting this error, I have the same "text" written on the text box.
Error occurred when executing CLIPSegDetectorProvider: object of type 'NoneType' has no len()
Did you install all requirements? The last requirement (6) won't automatically get installed by comfy Manager. You have to search for it and enable it yourself. You can however do it via the comfy manager as well, it's just not automatic.
Where do I find AestheticmixXL_v10.safesensors?
I've got an image with watermarks all over the place, but the mask it generated is only for the bottom of the image and a random line near the top, neither of which have any watermarks.
I tried playing with the mask_threshold and got the same result.
Is there anything else that needs to be done to pick up watermarks better?
I tried a second image, also 1024x1024 but different picture. Same result other than the line at the top in the image below -- the second one just masked the bottom part of this image as well.
It seems like even your example is just picking up stuff on the bottom of the image? What about watermarks all over the image itself?
It was mainly made to pick up the watermarks his model was gening near the edges but it was also picking up some false positives in the middle of the image from things like tattoos so I added an overlayed black image on top of whats detected to avoid that. If you bypass the maskmask section it might work for what you want but no promises.
@leinad882 Appreciate the info! Unfortunately, yeah, bypassing that just seems to throw a blur at the almost the whole image.
Might just try to work with some inpainting as I haven't seen anything great at detecting just watermarks.
Think if I inpaint and manually mask the watermarks myself, I could tap into the detailer section of this workflow?
@_POD_ If you do some manual masking very precisely over the watermarks and then pass it through the lama remover it might do well. I personally use this tool here for removing things but its kind of hard to install it free without the 1click installer they charge for(is possible to install it free though) GitHub - Sanster/IOPaint
I'm receiving this error on BBOX Dedector (SEGS) and i couldn't solve this. Do you know how to handle?
ValueError: Input image size (352*352) doesn't match model (224*224).Best
is this still working with comfyui current version? i get an error