Free background removal tool rembg tested: clean even with white shoes on a white background, but the default configuration has two pitfalls

rembg is an open-source background removal tool that can be used as a command-line tool, Python library, HTTP service, and Docker container. The repository uses the MIT license, while the model weights have separate licenses. X user ReignSwords recently ran a stress test with four images in a 4-core VM, using the u2net model, with images no larger than 800px.

All four images were tricky to cut out: curly-haired portrait 1.66 seconds, zoomed in twice to inspect the curly hair edges, clean; white shoes on a white background 1.66 seconds, separable even with zero contrast; tabby cat 2.86 seconds, whiskers intact; wine glass 2.70 seconds, with the bowl, stem, and base outlines all present.

Cold and warm inference were also tested: 1.42 seconds vs. 1.49 seconds, with no obvious warm-up effect. The portrait-specific model u2net_human_seg ran in 3.06 seconds, and its edges were no better than the general model, while requiring an extra 176MB download; the tester recommends skipping it.

But there are two pitfalls in installation and use.

First, pip install rembg does not give you the rembg command. The rembg i in.png out.png in many tutorials requires first installing the CLI extension, i.e., pip install "rembg[cli]". The repository now documents this in more detail: on a CPU environment, to get both the library and CLI you can use pip install "rembg[cpu,cli]".

Second, don't blindly follow the default model. rembg 2.0.86 defaults to bria-rmbg, which downloads about 1.02GB, and commercial use of the weights requires a paid agreement. If you want a small model, explicitly pass -m u2net, which downloads about 176MB and is Apache-licensed. On first use, the model is automatically downloaded to ~/.rembg/models.

Library usage is very short: from rembg import remove, open the image, call remove(), and get a PNG with an alpha channel. The alpha channel is the transparent background information. The tester checked the mask pixel by pixel.

There are also options for edge handling. For hard-edged products, cars, logos, and screenshots, the default mode is enough. If the background color is heavy and the cutout edges have colored halos, use -dc to decontaminate color fringes. If the shape of hair edges is wrong, then consider -a or -vm; -a is slow and may fail to converge on some images, while -vm downloads about 110MB more but can recover more fine hair strands. When batch processing and you can't check each image individually, -dc is more reliable.

Conclusion: for opaque objects, rembg is basically what people call "install and use." But the default configuration should not be accepted blindly; check the model license before commercial use.

Test data and conclusions come from ReignSwords' post on X; tool installation and model information come from the danielgatis/rembg repository. Original post: link, repository: link.