Background Remover
Cut the subject out of a photo and get a transparent PNG back. An open-source AI model does the work, and it runs in your browser — the picture is never uploaded. The model is about 26 MB, downloads once, and is ready whenever you come back.
Private by design: Your photo is processed by an AI model running in your browser and is never uploaded. The model itself is downloaded once from Hugging Face as ordinary site code and cached on your device. See how in-browser processing works.
How to remove a background
Drop a JPG, PNG, or WebP onto the upload area, or click to browse.
Press Remove background. The first time, the model downloads — about 26 MB, with a progress bar — and your browser keeps it for next time.
Check the result against the checkerboard, then download the transparent PNG.
What's actually happening
This tool uses MODNet, an open-source matting model released under the Apache 2.0 licence. It predicts, for every pixel, how much belongs to the subject and how much to the background. That prediction becomes the alpha channel of the output, so edges get soft partial transparency rather than a hard cut — which is what makes hair and clothing edges look right on a new background.
Everything runs on your device. Where your browser supports WebGPU the model runs on your graphics chip and a photo takes well under a second once loaded; otherwise it runs on the CPU and takes a few seconds. The model and the small runtime that executes it are downloaded from Hugging Face and jsDelivr as ordinary site files. Your photo is read into memory, processed, and never sent anywhere — you can watch the network tab and confirm no upload happens.
What it's good at, and what it isn't
MODNet was built for portrait matting — photos of people. That is what it is trained on and where it is reliable: a person against almost any background, including busy ones, with good handling of hair.
It often does a reasonable job on other clear, single subjects too — we tested it on an object against a plain background and it separated cleanly — but that is not what it was designed for, and results on products, animals, or cluttered scenes will vary. If your subject isn't a person, try it; just check the edges before you rely on it.
It removes the *background*. It does not let you point at one object among several and keep only that.
Why this model, and not a bigger one
We tested a larger general-purpose model (BiRefNet) first. It produces excellent masks on a server, but in a browser it could not run on ordinary hardware: its graphics-chip version needs more shader resources than typical GPUs allow, and its CPU version exhausts the browser's memory. Shipping it would have meant a 110 MB download that then failed for most people.
MODNet is a quarter of the size, runs on the hardware people actually have, and does the job it was designed for well. We would rather offer a tool that works for portraits than one that promises everything and fails.
The download, and why it happens once
The model is about 26 MB on a computer with WebGPU, or 12 MB on the CPU path — larger than an ordinary web page, and we won't pretend otherwise; it's the price of running a real neural network locally instead of on a server. It downloads the first time you press Remove background, with a progress bar, and your browser stores it in its cache. Every later photo, and every later visit, starts immediately. Clearing your browser's site data removes the cached model, and the next run downloads it again.
Good to know
The output is always a PNG, because JPG has no transparency. If you need a small file afterwards, run it through the Image Compressor.
One photo at a time. The model stays loaded between photos, so the second one is faster than the first — under a second on WebGPU.
Best results on photos of people. Other subjects can work but aren't guaranteed; check the edges.
Nothing is uploaded. The model file is fetched from Hugging Face's servers as ordinary site code — that is the only network traffic, and it contains no data from you.
Frequently asked questions
Is my photo uploaded to remove the background?
No. The AI model runs inside your browser, on your own device. The only thing downloaded is the model itself, from Hugging Face, and it contains nothing of yours. Your photo never leaves your computer — you can verify that in your browser's network tab.
Why is the first run slower than the rest?
Because the model — about 26 MB — downloads before it can run. Your browser then keeps it, so the next photo and every later visit skip that step. Once loaded, a photo takes well under a second with WebGPU, or a few seconds on the CPU.
Does it work on things other than people?
Sometimes. MODNet is trained for portraits, and that is where it is reliable. It can separate other clear, single subjects from a plain background — it did in our tests — but products, animals, and busy scenes are not what it was built for, so check the result before relying on it.
Why are the edges soft instead of sharp?
Because the model predicts partial transparency along edges rather than a hard cut. That is deliberate: it is what makes hair and fabric look natural when the cut-out is placed on a new background. A hard edge would look pasted on.
What is WebGPU, and does it matter?
WebGPU lets the model run on your graphics chip. Current Chrome and Edge on computers have it, as does recent Safari. Without it the tool still works on the CPU, just a few seconds slower per photo — you don't need to do anything.
Which model is this, and can I trust it?
It's MODNet, an academic portrait-matting model published under the Apache 2.0 licence with open code and weights. We chose it because it runs reliably on ordinary hardware and its licence permits use on a free, ad-supported site — several popular alternatives are non-commercial only, and a larger general-purpose model we tested could not run in the browser at all.