A Yolov8 detection model that detects watermarks in images.
The model can be used as an ADetailer model (for Automatic1111 / Stable Diffusion use), or using other inference scripts to return detection bounding boxes of watermarks.
The model is entirely trained on the following dataset: MFW-feoki/W6-janF
A small tutorial on how to use the model can be found on this Github: https://github.com/MNeMoNiCuZ/yolov8-scripts or this CivitAI article.
Description
FAQ
Comments (8)
Wow, this looks cool, was wondering if someone was going to be able to create this. Thanks.
I think it needs way more training and a broader dataset. But it's absolutely possible.
It should ideally be a segmentation model though, to get a better mask of the region, rather than a bounding box. This won't work well with watermarks that are diagonal across the image for example. The bounding box would cover the entire image essentially.
It's more reasonable for a community project though, a lot of work would have to go into it to create enough data for it. And I don't know the legality of training on watermarked images.
This is more a proof of concept that works for some cases.
this page says:
Important
Requires Python 3.11 or newer.
idk bout y'all but my gradio says py 3.10.6, will this matter?
Hmm, your gradio? Are you talking about Automatic1111?
To use the models in A1111 or Comfy you can use any version of Python. 3.10.6 is fine.
The documentation on the github page is if you want to train your own models. I think you may need 3.11 for that.
I hope that information helps @anon2071
Doesn't work that much on white watermark. Idon't know how did that work on the third picture
The images in my example image are from the same source as the training set. The images weren't used in the training, but they are "similar". So the model would work well with it, but not necessarily all styles of watermarks.
You could use this model as a base and add additional training on top of it to match the style of the watermarks you're trying to detect.
Is there a way to use this to remove watermarks? I'm getting detection but it just details the watermark
Yeah, in ComfyUI you can do something else with the detected area rather than "detail" it.
You can use something like Lama Cleaner to do it automatically with this model.
https://huggingface.co/spaces/Sanster/Lama-Cleaner-lama