How background removal works entirely in your browser (and why it matters for your photos)
Yesterday's news — remove.bg closing its standalone site on 1 December 2026 — sent a lot of people looking for a replacement. Most of the lists you'll find point at another upload-and-wait service. But there is a different kind of replacement, and it's worth understanding: a background remover that runs entirely in your browser, where your photo never leaves your device.
I built one (it's part of ToolNest), and building it taught me how the whole thing actually works under the hood. Here's the honest version — what the browser is really doing, where it struggles, and when you should use something else instead.
What "in-browser" actually means
A normal online background remover works like this: you upload a photo to a server, a model on that server finds the subject, and you download the result. Your photo travels to a stranger's infrastructure and, depending on their retention policy, may sit there for hours or days.
A browser-based remover flips the pipeline around. The model — a small segmentation network, usually a U-Net-style architecture trained to separate foreground from background — is downloaded once as a compact file (a few megabytes, often running through ONNX Runtime Web or TensorFlow.js) and then executes locally, in your tab, using WebAssembly or WebGL. Your photo goes from your disk, into your browser's memory, through the model, and back out as a PNG with transparency. Nothing is uploaded, because there is nothing to upload to.
That distinction is the whole product. Speed is a bonus; privacy is the point.
The pipeline, step by step
- Load and downscale. The model doesn't need your 24-megapixel photo. It works on a downscaled copy (typically 320×320 or 512×512) — enough resolution to find edges, small enough to run in milliseconds.
- Segment. The model outputs a mask: for every pixel, a probability that it belongs to the foreground. This is the expensive step, and it's the one that used to need a GPU server. Modern browsers do it on your CPU or integrated GPU in under a second for a typical photo.
- Refine the mask. A raw mask has soft, noisy edges. A quick refinement pass — feathering, morphological cleanup, sometimes a guided filter using the original image — turns a blobby outline into something that looks cut out rather than torn out.
- Composite. The refined mask becomes the alpha channel of your original full-resolution photo. The subject keeps every original pixel; only the background becomes transparent.
Where it struggles (the honest limits)
- Hair, fur and translucent edges. This is where every remover — free or paid, browser or server — fails first. A mask at 320×320 simply cannot resolve individual hairs. Good tools feather the edge so it looks natural at normal zoom; great tools let you paint the mask back in by hand.
- Busy backgrounds. If the background has the same colours and textures as the subject, the model has to guess. A clean, well-lit photo with a contrasting background gives the best result by a wide margin — which is also why product photographers use plain backdrops.
- One photo at a time. The browser is doing real work per image. If you need 200 product shots processed in an afternoon, a scripted local workflow (open-source rembg on your own machine, for example) is the right tool — and it's free too.
- Big files, slow devices. A 50MB photo on a five-year-old phone will take its time. Downscaling helps, but there's no escaping the compute.
When to use which kind
- One photo, personal or client work: browser-based. No upload, no account, no credits, no retention policy to read.
- A batch of product shots: local script (rembg) or a desktop app. Same privacy, far more throughput.
- A hard edge case (wedding hair, glass, fur): a desktop editor with manual masking, or a paid service — and check what happens to your file afterwards.
Why I built it this way
ToolNest is 144 free tools, and every one of them runs client-side. That started as an architecture preference and became the product's entire promise: your salary after tax, your weight, your client's product photos are not a website's business. Close the tab and the numbers are gone, because they were never anywhere else.
remove.bg closing is a reminder of why that matters: when a free service is someone else's server, it can change price, change terms, or close on three weeks' notice. A tool that runs on your own machine can't be taken away on a deadline.
The background remover is at https://toolnest-xi-silk.vercel.app/tools/background-remover/ — open it, drop a photo in, download the result. Try it on your hardest image first, zoom in on the edges, and tell me where it fails. That's how all 144 tools got built, one demand at a time.
