Back in 2020, I wrote about how I removed the backgrounds from 11,782 JPGs for free using AI.
At the time, the solution involved Python, TensorFlow, U-2-Net, Google Colab, Jupyter Notebook and a collection of scripts I modified for the project. It worked—and it did about 95% of the heavy lifting—but it was not exactly a user-friendly process.
I had to upload the images in batches of around 1,000, connect Colab to Google Drive, run the scripts, download the results and reset everything before starting the next batch.
The entire project took me approximately two weeks to research, build and refine.
It was free, but it definitely wasn’t convenient.
Six years later, I ran into the same problem again.
A New Project, 600 More Images
Today, we started working on a project that required us to update approximately 600 JPG images.
Once again, the backgrounds needed to be removed and the finished images needed to be saved as transparent PNGs. The original files also had to remain untouched, and because the images were organized into folders and subfolders, the output needed to preserve that structure.
I could have gone back to the old Colab workflow. I could also have uploaded the files to one of the many online background-removal services that now exist.
Instead, I thought: there should be a simple desktop application for this.
So I built one.
Introducing Drive Background Remover
Drive Background Remover is a free Windows utility designed to remove image backgrounds in bulk.
You select individual images or an entire folder, choose an output location and let the application process the batch. Each finished image is saved as a transparent PNG without changing or overwriting the original.
The application can:
- Process JPG, JPEG, PNG, WebP and BMP images
- Add entire folders at once
- Optionally scan subfolders
- Recreate the original folder structure in the output directory
- Preserve the original filenames
- Avoid overwriting existing output files
- Process images entirely on the local computer
- Produce transparent PNG files automatically
There are also multiple quality settings. The included Fast model works immediately, while larger Balanced and High Quality models can be downloaded when needed.
It Doesn’t Use an LLM to Process Images
Because everything is suddenly being labelled “AI,” it is worth explaining what the application actually does.
Drive Background Remover does not send images to ChatGPT or another large language model. It uses computer-vision models designed specifically to separate the subject of an image from its background.
The application is built with Python, rembg, ONNX Runtime and PySide6. The image-processing models run directly on the user’s computer.
That means the images are not sent to Drive Marketing, an LLM provider or a cloud image-processing API. Once the required model is available, processing can happen offline.
For client and product photography, keeping the files local is a meaningful advantage.
Two Weeks Versus One Afternoon
The original project took me approximately two weeks to complete.
The Windows application described in this article took an afternoon—and I was working on other parts of the client project between prompts.
That difference is difficult to overstate.
Tools such as ChatGPT and Claude Code have fundamentally changed the way I approach software development. They allow me to move from an idea to a working implementation much faster, especially when a project involves unfamiliar libraries, packaging systems or operating-system-specific behaviour.
However, I would not describe this project as “vibe coded,” at least not in the way that term is normally used.
I did not generate a pile of code, assume it worked and publish it.
The project was connected to a proper GitHub repository from the beginning. I reviewed the code, tested the application, diagnosed errors, made architectural decisions and hand-coded changes when necessary. We worked through Windows security restrictions, model dependencies, filename handling, nested folder structures, interface design, branding, packaging and installer behaviour.
The tool was repeatedly rebuilt and tested as those decisions were made.
The AI accelerated the implementation, but it did not take responsibility for the result. That still belonged to me.
Coding Is Only Part of Development
This project also demonstrates why these tools can be especially powerful in the hands of traditional developers.
Writing code is probably only about 30% of actual software development.
The rest involves understanding the problem, defining requirements, choosing an architecture, evaluating dependencies, anticipating edge cases, reviewing output, testing behaviour, making security decisions and deciding what the software should—and should not—do.
For example, removing a background from one image was never the real problem.
The real problem was building a workflow that could safely process hundreds of files without overwriting the originals, preserve filenames, mirror subfolders, handle output collisions, provide different quality levels, work offline and remain understandable to someone who had never used Python.
Those are product and engineering decisions. Generating code is only one part of implementing them.
A developer who already understands those responsibilities can use AI tools as an extremely effective accelerator. You can recognize when a suggestion is wrong, identify missing requirements, review the implementation and redirect the work before a small mistake becomes a structural problem.
AI does not eliminate the value of development experience. In projects like this, it amplifies it.
From a Notebook to a Windows Application
The other interesting part of revisiting this problem was seeing how much the underlying technology has improved.
In 2020, I was learning how Jupyter Notebook, Google Colab, TensorFlow and U-2-Net fit together. The final workflow worked, but it was essentially a technical experiment adapted to solve a real business problem.
This time, I was able to turn the same basic idea into a normal Windows application.
There are no notebooks to configure, no code blocks to run and no Google Drive folders to reconnect. The user does not need Python installed and does not need to understand how the model works.
Download the installer, launch the application, select the images and choose where the transparent PNGs should go.
That is the entire workflow.
What About the Results?
As I noted in the original article, automated background removal is not perfect.
Results depend heavily on the image. A clearly defined product against a contrasting background will usually perform better than transparent objects, fine hair, shadows or a subject whose colours are similar to the background.
For a large batch, I still recommend reviewing the finished images and manually correcting any important files that contain artifacts.
The goal is not necessarily to eliminate every second of manual work. It is to automate the repetitive 90–95% so that a designer only has to concentrate on exceptions.
For hundreds—or thousands—of images, that is a substantial difference.
Download Drive Background Remover
Drive Background Remover is currently available for Windows 11.
The application runs locally, does not require an account and leaves the original images untouched.
Download Drive Background Remover for Windows
Six years ago, solving this problem took approximately two weeks, a collection of scripts and a slightly awkward Google Colab workflow.
Today, I turned it into a packaged Windows application in an afternoon—while still working on the client project that created the need for it.
That feels like progress.
