withoutBG: Open-source local background removal for GIMP, macOS, and Docker

Software GIMP
1

Hi everyone,

I’ve been working on withoutBG, an open-source project for AI background removal.

There are a few ways to use it:

  • GIMP 3 plugin that integrates directly into GIMP and creates an editable layer mask instead of permanently deleting the background.
  • Native macOS app with drag-and-drop support and a built-in Local API. The GIMP plugin can connect to it, so you don’t need Docker on macOS.
  • Docker image for Linux, Windows, self hosting, or automation.

withoutbg-gimp
withoutbg-gimp1211×907 817 KB

withoutbg-mac
withoutbg-mac2346×1352 1010 KB

Everything uses the same inference engine, so you can choose the workflow that fits your setup.

I would really appreciate feedback from the PIXLS.US community, especially on installation, workflow, performance, image quality, and support for difficult images. Feature requests and bug reports are always welcome.

GitHub

Thanks for taking a look. I am happy to answer any questions or discuss the implementation.

3 Likes
2

Thanks you for developing this, I will have a look! Couple questions…

  1. Apple provides a background removal tool right out the box. Is different? Can it handle lager image?
  2. Is it also possible to export the mask itself?
  3. Does the Gimp plugin need the api?
3
  1. withoutBG generally delivers better results than the tool in macOS. You can test and see. Some results: Pro Model Results — Alpha Matte Quality Examples
  2. Yes. The local API is actually returning the alpha matte, not image without background.
  3. Yes, it does. Because I can’t bundle such a big ML model into a GIMP plugin.
4

How big is it?

5

I asked the same question on imrans post on g-f-dot-net and it is;

withoutbg-openweights-v3-service-cpu is the lightest Docker image and is about 927 MB, including the model and runtime.

I am trying out a different “lighter” AI bg-removal Gimp plugin, a local installation with “only 600MB” of dependencies. As with most things image editing, it all depends on the image. Ai can work very well or it can leave you with a bit of manual editing.

6

Thanks for responding the above question and your feedback.

There is definitely a tradeoff between quality, speed, and model size. I think withoutBG Open Weights is a good balance for desktop use.

For this release, Docker was the most straightforward and hassle-free solution. However, I plan to improve this further. I’m considering building an easy-to-install standalone CLI tool that can also act as an HTTP server, which should significantly reduce the installation requirements.

Regarding the model size, there are lighter matting and segmentation models available, but they generally struggle more with transparency and fine details such as hair, fur, and semi-transparent objects.