Is it possible to bring this background remove AI to GIMP?

kevin_tee · December 23, 2018, 4:24pm

So I found out about this

I automatically detect face and remove the background it.
Which use this algorithm (I think)

More info: This free online tool uses AI to quickly remove the background from images - The Verge

paperdigits · December 23, 2018, 5:29pm

Technically it is possible, but it probably won’t happen as the necessary computer power is too high for most people’s machines to run the software adequately.

imran · July 29, 2026, 12:58pm

Yes. I recently built an open-source plugin for GIMP 3 that runs AI background removal locally. It creates a transparent layer mask, so you can continue editing it with GIMP or place a new background underneath.

GitHub: GitHub - withoutbg/withoutbg-gimp: GIMP 3.0 plugin for local WithoutBG background removal · GitHub

Installation and documentation: GIMP 3 Plugin for Background Removal

Pete.IRL · July 29, 2026, 5:36pm

I’m not so positive abot BG removers, but then again I haven’t used Ai or browser based. Is it common thar they leave Halo effect or other artifacts which then need further work…generally manual and not so quick.

Would they work better if the background is made birighter or darker before BG removal…what is the best way to do that, or isolate a background.

As I’m at it, a curious question popped into my mind earlier. When shooting raw, does the camera/software recognise foreground from that which is more distant and can they be seperated in processing.

afre · July 29, 2026, 6:47pm

Removed spam before Pete’s post. Why would someone necro an eight year old thread as well as another old one?

@Pete.IRL I believe that some of our post-processing software can access focus point data in certain raw files.

paperdigits · July 29, 2026, 8:25pm

Because its on topic and answers the question posed?

imran · July 29, 2026, 8:34pm

When shooting raw, does the camera/software recognise foreground

There isn’t a hardware solution for this in most cameras. Some cameras and smartphones have depth sensors, which can be used to blur the background.

The actual foreground/background separation is generally a software problem. Today, the most powerful solutions are based on deep learning.

which then need further work

Correct. It is a difficult problem. Unfortunately, my previous reply was deleted because it was identified as an advertisement, but the plugin I posted (withoutBG) simply generates an alpha matte without applying it directly as an alpha channel. This gives users the freedom to refine the mask manually in GIMP, while still speeding up the matting process.

afre · July 29, 2026, 8:59pm

Sorry for the trigger finger folks. (Edit: undeleted by Mica and now made visible by me).

@imran Could you explain what “Meta DINOv3 License for DINOv3 backbone weights” means for FOSS users? Thanks.

imran · July 29, 2026, 9:26pm

Sure. First, withoutBG is an open weights model. The training code is not open source, but the trained withoutBG model weights are released under the Apache 2.0 license. The withoutBG GIMP plugin and the withoutBG inference library are also open source under Apache 2.0.

Regarding the Meta DINOv3 License: withoutBG uses the DINOv3 backbone as a feature extractor. It extracts high level image features that help the matting model identify objects, fine details such as hair, and transparent or semi-transparent regions.

The only component that is not Apache 2.0 is the DINOv3 backbone weights, which are distributed under Meta’s DINOv3 License. Therefore, anyone using withoutBG with the included DINOv3 weights must also comply with that license.

In short:

withoutBG model weights: Apache 2.0
withoutBG code (plugin and inference library): Apache 2.0
DINOv3 backbone weights: Meta DINOv3 License

paperdigits · July 29, 2026, 9:33pm

can you link us to that license? Is that part optional?

imran · July 29, 2026, 9:39pm

The DINOv3 license is here:

github.com/facebookresearch/dinov3

LICENSE.md

main

# DINOv3 License

*Last Updated: August 19, 2025*

**“Agreement”** means the terms and conditions for use, reproduction, distribution and modification of the DINO Materials set forth herein.

**“DINO Materials”** means, collectively, Documentation and the models, software and algorithms, including machine-learning model code, trained model weights, inference-enabling code, training-enabling code, fine-tuning enabling code, and other elements of the foregoing distributed by Meta and made available under this Agreement.

**“Documentation”** means the specifications, manuals and documentation accompanying
DINO Materials distributed by Meta.

**“Licensee”** or **“you”** means you, or your employer or any other person or entity (if you are entering into this Agreement on such person or entity’s behalf), of the age required under applicable laws, rules or regulations to provide legal consent and that has legal authority to bind your employer or such other person or entity if you are entering in this Agreement on their behalf.

**“Meta”** or **“we”** means Meta Platforms Ireland Limited (if you are located in or, if you are an entity, your principal place of business is in the EEA or Switzerland) or Meta Platforms, Inc. (if you are located outside of the EEA or Switzerland).

**“Sanctions”** means any economic or trade sanctions or restrictions administered or enforced by the United States (including the Office of Foreign Assets Control of the U.S. Department of the Treasury (“OFAC”), the U.S. Department of State and the U.S. Department of Commerce), the United Nations, the European Union, or the United Kingdom.

**“Trade Controls”** means any of the following: Sanctions and applicable export and import controls.

By clicking “I Accept” below or by using or distributing any portion or element of the DINO Materials, you agree to be bound by this Agreement.

This file has been truncated. show original

In the current release, it isn’t optional. Both the inference Docker image and the macOS app include the DINOv3 backbone, so they also include the DINOv3 weights:

My own code and the withoutBG model weights are licensed under Apache 2.0. The DINOv3 backbone weights retain Meta’s DINOv3 license.

Pete.IRL · July 29, 2026, 11:21pm

I’m a bit lost here. I downloaded the (remove_bg_gimp3.py) from github.

I placed the file in a folder named (remove_bg_gimp3)

I moved the folder to (C)

I went to Gimp > Edit > Folders.

I see Temp folder (gimp 3.2…) & Swap folder options

What do I do next?

imran · July 30, 2026, 7:34am

Can you follow this guide: GIMP 3 Plugin for Background Removal

Tobias · July 30, 2026, 9:02am

If you are interested in camera-generated depth maps, I have written some blog posts on the topic. Most modern phones create depth maps and embed them in photos as metadata. Unfortunately, GIMP and darktable can’t extract this data, but you can usually do so with ExifTool.

And if someone is interested in fixing it in GIMP here is the issue:

imran · July 30, 2026, 7:06pm

Thanks for the creating the issue. I needed a few times and it was hard to extract with popular image processing tools.