remove.bg is shutting down: 7 free local background removers, license-checked

What's going on

On December 1, 2026, at 9 a.m. CET, remove.bg is history. According to the official FAQ, which several sites quote word for word, the website will no longer be reachable after that. The remove.bg API page says the self-serve API is moving to Leonardo.Ai, which, like remove.bg, belongs to Canva. Unused pay-as-you-go credits expire the same day, and the terms of service say there's no refund. Annual subscriptions are supposed to get a pro-rated refund. Download your invoices and usage reports from your account before the deadline, because you won't be able to afterwards. According to the FAQ, enterprise contracts aren't affected.

We asked a simple question: what can you run for free on your own machine, and what are you allowed to do with it commercially? One important caveat up front: we didn't run any measurements ourselves. We analyzed published benchmarks, independent hands-on tests and the models' license texts, and we checked the math. Where numbers come only from the vendor, we say so.

What the move will cost you

If you stay, you either pay more or get less. Leonardo.Ai's migration guide lists $0.1047 per image. For 1,000 product photos a month, that's about $105. For comparison, a post on dev.to puts remove.bg at around $9 a month for 40 images, or a little over 22 cents per image. So Leonardo is cheaper per image. But you'll need a new account and a new key, and you'll have to rework your code: authentication uses a bearer token instead of the X-Api-Key header, and the response comes back as JSON instead of an image file. Results are saved to your Leonardo library by default unless you set ephemeral: true.

If you used the website, you'll end up at Canva. According to Canva's help pages, background removal in the editor is limited to photos under 9 MB. Images over 10 megapixels are downscaled after the background is removed. Do the math and a 24-megapixel photo (6000×4000) comes back at roughly 3870×2580 pixels. Several sources also describe the tool as a feature of the paid plans.

The contenders and their licenses

Almost every serious open background remover is built on BiRefNet, a model for "dichotomous image segmentation": every pixel is assigned either to the subject or to the background, even at high resolution. Most variants still produce a soft alpha mask, a transparency map with in-between values.

Three things stand out when you read the licenses.

The rembg trap. The popular command-line tool rembg is itself MIT-licensed. But since a commit on August 17, 2026 ("make bria-rmbg the default model"), its default model has been RMBG-2.0. The source code reads new_session(model_name: str = "bria-rmbg"). So if you just type rembg i foto.jpg foto.png, you're using a model that its Model Card says is released only "for non-commercial use". The README points this out itself, but older tutorials and tests still list u2net as the default. If you run a shop, always pass -m birefnet-general or -m isnet-general-use. rembg has also picked up two cloud backends (withoutbg and atlascloud). Choose either one and your images go back out over the internet.

withoutbg isn't "just Apache". The Model Card names its own "withoutBG Open Weights" license. The components taken from Meta's DINOv3, a pretrained vision model that serves as the backbone here, fall under the DINOv3 license. That license doesn't rule out commercial use. But among other things, it prohibits military uses, requires you to include the license when you redistribute the model, and lets Meta terminate it if you violate the terms. For developers building the model into their own products, that's a lot more paperwork than MIT. The edge refiner is only available through the cloud API.

The training-data gray zone. Even MIT-licensed weights were often trained on datasets released for research only. The Lucida README is upfront about this for DIS5K, COD10K and P3M-10k, and says: "Commercial users should make their own assessment." Whether that legally "taints" the weights is unresolved. We're not lawyers.

What the tests say about hair, glass, spokes and text

The only public benchmark with separate categories comes from Lucida, an MIT-licensed fine-tune of BiRefNet-HR, meaning a BiRefNet retrained for hard cases. It measures the mean error against a hand-made alpha mask (MAE, lower is better) across 203 images. Note that the benchmark was created by Lucida's own author.

By our math, the freely licensed BiRefNet-HR comes out about 14 percent ahead of the non-commercial RMBG-2.0 overall. On hair, the two are practically tied.

Hair: The hands-on tests disagree here. A ComfyUI test on Medium put BiRefNet-HR on top, though the author prefers BEN2. Invideo names BEN2 the leader, while Fuser praises RMBG-2.0 on backlit curls. A CPU test on dev.to using the small BiRefNet variant (Swin-Tiny) lost fine individual hairs and left blue color fringes from the old background. rembg now has a -dc option that cleans up fringes like these.

Glass: This is every model's weak spot. Fuser found that RMBG-2.0 keeps a perfume bottle intact but treats the glass as a solid object: the café background still shows through when you drop it onto a new one. In aireiter's test, an iced-tea glass survived with BiRefNet and ISNet, while u2net erased both the glass and the hand. So "kept" doesn't mean "properly transparent".

Spokes and grilles: A public issue in the browser project nukebg uses motorcycle spokes, of all things, as its test case. There, the small INT8 exports of BiRefNet-lite, versions shrunk down to 8-bit integers, show an unresolved "sawtooth" artifact. ISNet and MVANet still haven't been tested.

Products: With u2net on the CPU, a shoe came out cleanly in the dev.to test, after 8.2 seconds at full resolution. In Fuser's test, on the other hand, rembg completely lost a bottle's gold cap.

Text screenshots: We didn't find a single test. Our take: for screenshots with a solid-color background, a classic magic wand tool is usually more reliable than a segmentation model trained on photos.

FeyNoBg: strong on paper, not independently verified

FeyNoBg from Feyn came out at the end of July 2026. It's a BiRefNet with a deeper third stage (24 instead of 18 blocks, 263M instead of 222M parameters), released under Apache 2.0. According to the maker, it reaches an S-measure (a score for how faithfully the mask follows the shape) of 0.961 on DIS5K and 0.981 on UHRSD-TE, for example. On CAMO-TE, it trails BiRefNet by 1.9 percent. However, the comparison is against the "best published" numbers from other papers, not a consistent run of its own. Feyn itself admits that the matting data was converted into hard black-and-white masks. So the model never learned soft transparency values, which can be a problem for glass and hair. Kompozy, a competitor that discloses this, rates FeyNoBg as "the best matting model to self-host, but not a finished product". There's no command line, only Python via the nobg package. The ComfyUI-RMBG plugin has supported FeyNoBg since version 3.3.0, released October 20, 2026.

Speed: CPU or GPU

• GPU: According to its developer, BiRefNet needs about 3.5 GB of VRAM at 1024×1024 pixels in FP16 and 57.7 ms per image on an RTX 4090. Feyn measures about 3 GB of peak usage and 878 ms per image for its library, but doesn't say what hardware it used.
• CPU: BiRefNet Swin-Tiny averaged 4.3 seconds on an M1 Max. BiRefNet-general via rembg took 15.2 seconds in aireiter's test, ISNet 1.1 seconds and u2net 0.4 seconds. By our math, 40 product photos with BiRefNet on the CPU take about 10 minutes. That's fine for occasional batches, but not for thousands of images.

The tools built into your operating system

On macOS, you can remove backgrounds in Preview via "Tools → Remove Background", with a Quick Action in Finder, and with "Lift Subject", using on-device machine learning according to MacMost. Windows 11 offers one-click background removal in Paint and the Photos app. Our sources don't say whether processing happens locally. Samsung Galaxy devices have Image Clipper. Google Photos for Android has been creating stickers since early 2026, reportedly using on-device machine learning on devices with at least 4 GB of RAM. All of these tools are free and make it easy to cut out single images. What they mostly lack is batch processing, resolution control and fine-tuning of edges. According to MacMost, Preview often grabs a bit too much background.

Our verdict

Our take: for most shops, BiRefNet is the obvious replacement. It's freely licensed, easy to reach through rembg, and in the only category benchmark it beats RMBG-2.0 overall. The biggest risk isn't quality, it's convenience: if you call rembg without -m, you've been using a non-commercial model since August. It remains to be seen whether FeyNoBg's vendor numbers hold up in independent tests. And to be honest, no model has solved the glass problem. Bottles and glasses still need manual work, whether you run locally or in the cloud.