Seamless Image Background Removal with withoutbg

Introduction

Whether you work in professions which work with images a lot (graphic designers, e-commerce sellers), professions with medium usage of images (educators, bloggers, small business owners) or hobbyists (photgraphers, craft enthusiasts, amateur content creators), image editing tools might have come across your personal software list.

Although there are some great software tools out there that can satisfy such features fairly well, most of them aren't free. In addition, even fewer of such software products are open-source and locally executable to protect your products or data.

In order to provide you with a good alternative for this use case, we want to showcase withoutbg, a tool freely available for download via pip for python.

Why use withoutbg?

In addition to costs and local executability, we want to highly a few problems well known tools might have:

1. Data privacy: Any tool that processes your imagery products or data raises the question of how that data is used.
2. Copyright and Intellectual Property: Some companies or tools might require you to drop intellectual property rights when processing your images.
3. Image Distortion: When using services where you cannot understand how they work, they might distort your images in a way that might be harmful or misinforming. This is particularly the case when using services that rely on neural networks to provide their services.

Remark: While withoutbg is a tool that uses neural networks to provide its functionality, we believe in the potential of this technology for this use case. While AI tools can provide harmful results, we have not encountered such an instance in our experiments. We firmly advocate for responsible and ethical usage, understanding the importance of thoughtfulness and exploration when employing such potentially intransparent tools.

Let's begin a rundown of software, that is popularly used for image editing and background removal:

1. Adobe Express: With Adobe being a dominant player in creativity tools (Source), it eventually developed Adobe Express, a tool for fast and streamlined image editing processes. While the tool itself is reviewed to be a great asset for simple imagery, it lacks in high-quality batch processing features and personalization (Source1, Source2, Source3). Alongside this, Adobe has faced severe criticism for integrating AI training into its software licensing (Source1, Source2). Even worse, Adobe has severe issues when it comes to collecting user data - a list of issues can be reviewed at tosdr.org. Finally, the usage of this software decorates the product with a price tag of currently ~12€/month per single user, lowered to ~4€/month per user for businesses. (Date of Research: 12.10.2025)
2. Clipping Magic or PicsArt: While performing a bit worse than Adobe Express, Clipping Magic or PicsArt are two popular tools to provide similar functionality. While Clipping Magic had several users report an overly strict refund policy with bad business practices (Source). The same can be added about PicsArt, whereas users have reported PicsArt to overcharge users drastically. (Source)
3. Gimp or Krita: Both highly rated software tools when it comes to hobbyist creative suites, both are capable of in-depth image editing and manipulation. However, their learning curve can be too much for newcomers in the field. Some users reports indicate learning the user interface alone for Gimp can take 2-4 weeks. (Source)
4. Other online tools like Photopea: While being functional in the way this tool is being used, users have reported hefty amounts of advertisements. This behavior can diminish productivity, leading some users to leave the platform altogether.

While both Gimp and Krita are recommendable tools for this task, some users might not have the time nor the capacity to develop skills for creative suites. For those users, we want to provide insights into withoutbg.

Installation and Usage

As per the GitHub page of withoutbg, we can download withoutbg seamlessly in a python environment like anaconda or alternative package managers.

pip install withoutbg

Now, we can navigate into a directory which holds an image - in this instance with the name of test.jpeg. Nextup, we can run the withoutbg command to remove the background.

withoutbg test.jpeg

After the initial download of the neural network which provides the background removal, the image is automatically being processed. We can then find an image with the name of test-withoutbg.jpeg in our original directory, which holds the processed results.

If we want such images to be png's - because we want to utilize the transparent background - we can take advantage of python in our environment and run the following python script.

from PIL import Image

import os

files = [f for f in os.listdir('.') if os.path.isfile(f)]
for file in files:
    if ("withoutbg.jpeg" in file):
        image = Image.open(file)
        image.save(file[0 : file.find(".")] + ".png")

Examples

With our initial setup concluded, we want to showcase some of our experiments to highlight the quality and restraints of the software. We begin by experimenting with clear images showcasing animals, which have easily identifiable backgrounds.

From these images, we can see that the processing of the images provides high-quality results. Despite the fact that these changes could be done by a professional digital artist in a few minutes, this tool automates such a task for amateurs in this field.

Next up, we jump into processing of trees. The idea is to track the branches here and see if they are being cropped in any way.

Our observations of these images show us, that the branches are being handled perfectly. The top left image is a really good example for this, where we can see that two branches are hanging down from the tree - they still remain within the processed image.

As a next step, we step up the difficulty. Now we utilize images presenting technical products, which show a lot of details. The idea here is to understand how the model prioritizes details as "not background" and how they are being removed from the image.

The results here show some of the limitations of such an automated software tool. We see that the model cannot always properly decypher which objects should be considered foreground, which can be observed in the image containing the robot. We observe multiple background objects (like the bench) which are still being rendered. In the image about the tablet, we can still see a lot of objects around it, with some of them losing coherence. Practically nothing has been edited. Lastly, the laptop has lost coherence aswell, whereas this mostly focuses on the keyboard area.

Remark: Keep in mind that despite our "clear" descriptions, different observers might perceive different objects to be in the "foreground" of the images. Therefore, quality assessments have to be made on a case-by-case basis.

In the next two galleries, we wanted to showcase how groups of people are being processed.

Within the last part of our experiment series, we can observe that contrast can make a huge difference for the model. When observing the image about the two people walking down a shopping street, we can observe that some details in the background still remain, while both people are still unchanged. The image of two people seemingly being involved in an interview has been processed perfectly, aswell as the people in the fog. The other pictures suffer from either foreground objects being incorrectly processed or background details still being maintained despite the fact that they should not.

TL;DR

In this article, we introduced withoutbg, a freely available open-source tool for removing backgrounds from images using neural networks. We introduced the tool in the light of data privacy, cost reductions and intellectual property rights, focusing on commercially available tools in this context. Withoutbg can be easily installed using pip and python and used to automate background removal tasks for high-contrast images. While demonstrating
impressive results with clear images (animals, trees), the document also highlights limitations – the tool struggles with complex scenes and detailed technical products, sometimes misinterpreting foreground and background elements.