The thing that prompted it
remove.bg is the tool most people mean when they say “background remover”, and its standalone site closes on December 1, 2026; the function moves into Canva, and the API to Leonardo.Ai. That left a lot of people looking for somewhere to go, and the obvious replacements have the same shape as the thing they are replacing.
The shape is this: the free tier gives you a preview. remove.bg’s free download was capped at 0.25 MP (~625 × 400) — enough to see that the cutout worked, not enough to use. To get the file at the size you supplied, you paid per image or subscribed. That is a reasonable business when every cutout occupies a GPU somewhere. It is also, from the other side of the screen, the experience of doing the work twice: once to find out it worked, and once to pay for it.
The full comparison with remove.bg and Canva has the numbers, tier by tier.
What we did differently
We moved the model to your side of the network. The segmentation runs in your browser through WebAssembly — U²-Net small · 4.4 MB for the immediate result, IS-Net int8 · 42 MB behind it for better edges — and your image never travels anywhere. There is no upload, and there is no server here that could receive one.
Most of what people find good about the tool falls out of that single decision rather than from any generosity on our part. Nothing is metered, because the hundredth image costs us exactly what the first one did: nothing. Exports come out at full resolution, because there is no bandwidth bill to shrink them for. There is no account, because there is nothing to attach one to. The editor keeps working with the network off. And a photo under NDA, a medical image or a passport scan can go through it without a conversation about a retention policy, because the conversation is settled by the architecture: see the privacy policy for the details, and the network panel for the proof.
The costs of that choice are real and land on you rather than on us: roughly 60 MB of runtime and model files on the first visit, cached afterwards, and inference that runs at the speed of your laptop instead of a datacentre GPU. We think that is the right trade for most people most of the time. It is not the right trade for everyone.
What this tool is not
It will not beat a mask cut by hand in Photoshop. On flyaway hair against a busy background, on fur, on glass and on smoke, a retoucher with a pen tablet and an hour will produce something this cannot — that is not a gap we expect to close, and any automatic tool telling you otherwise is selling you a preview at 625 × 400. The refine brush is here because we assume you will sometimes need to fix the edge yourself.
It is also not a photo editor. There is no retouching, no colour grading, no layers. It removes a background, puts something else behind the subject, and exports the result at the size you need.
And it is not magic on old hardware. A large image on a five-year-old phone will be slow, because the work is genuinely happening there. The Fast model exists for exactly that case.
Who makes it
The service is operated by Aman Jaiswal (The Change BG).
If something is broken, or the cutout failed on an image you would expect it to handle, we would like to hear about it — get in touch.