IMG-009

Remove Image Background
Create a transparent cutout with a segmentation model that runs inside your browser. Your photo is never uploaded. The trade is time: the model itself is a 44 MB download the first time, and a cutout takes roughly 35 seconds on a desktop machine.
- Free
- No signup
- Private · runs locally
Background removal runs locally in your browser. On first use, the AI model is downloaded and cached by the browser.
How to remove a background
- Choose a JPG, PNG or WebP, up to 50 MB.
- Wait. On the machine we measured, the first run took about a minute because the model has to arrive first, and later runs took roughly 35 seconds.
- Check the transparent preview and download the PNG.
There is nothing to adjust, and that is deliberate. The model decides what the subject is, and the only input it takes is the picture.
Which model this is, and what it costs you
The cutout is made by IS-Net, a segmentation model published for dichotomous image segmentation, in the 8 bit quantised build that ships with the @imgly/background-removal library. It runs through the ONNX runtime compiled to WebAssembly, which is the part that makes a model like this work in a browser tab at all.
The model is not in this page. The first time you use the tool your browser fetches it, 44 MB of weights plus about 12 MB of runtime, from the library's own content network. That happens once and is then served from your browser's cache. Your photo is not part of that exchange and is never sent anywhere.
It runs on the processor rather than the graphics card, and in a single thread, because the multi threaded build needs cross origin isolation headers that would break other things on this page. That is the honest reason this takes seconds rather than being instant.
Why the time barely changes with image size
The model always sees a 1024 by 1024 version of your picture, whatever you give it. The mask it produces is then scaled back up to your original dimensions, and the file you download has exactly the pixel size you started with.
Measured on a desktop machine with the model already cached, a 400 by 300 image took 38 seconds and a 1200 by 900 image took 38 seconds. The work is fixed by the model's input size, not by your file, so a large photo is no slower than a small one. A phone will be slower than both.
The same fact explains the edges. On a 6000 pixel photograph the outline you get is a 1024 pixel outline stretched back out, which is why fine detail such as flyaway hair or fur comes back soft while a clean product outline comes back sharp.
Where it does badly
A single, clearly separated subject is what the model was trained for: a person, a product, an animal, an object on a contrasting surface. Give it that and the cutout is usually good enough to use as it is.
It struggles with several subjects at once, because it is looking for one salient thing and will often keep the largest and discard the rest. It struggles with glass, smoke, water and anything else that is meant to be partly transparent, because the mask is a decision about each pixel rather than a measurement of how transparent it is.
It also struggles with very wide or very tall pictures. The image is squeezed into a square before the model sees it, not letterboxed, so a panorama arrives at the model distorted and the mask comes back correspondingly worse. Crop a wide shot closer to the subject first and the result improves.
Why use this rather than a hosted remover
The hosted services are faster and several of them are more accurate, because they run larger models on their own hardware. They do that by taking a copy of your picture.
This one does not. For an identity document, a photograph of a child, an unreleased product shot or anything under an agreement, that difference matters more than the seconds. For a holiday snap it probably does not, and we would rather say so than pretend otherwise.
Frequently asked questions
Is my image uploaded anywhere?
No. The picture is read from your disk, processed in the page and handed back. The only thing that comes over the network is the model itself, once, and that request contains nothing of yours.
Which format is generated?
PNG with an alpha channel, at the same pixel dimensions as the image you put in. PNG is the only sensible choice here, because JPG has no way to store transparency at all.
Why is it slow the first time?
Because the model has to arrive before it can run: 44 MB of weights and about 12 MB of runtime. After that your browser serves both from cache and the wait drops to the processing time alone, which measured roughly 35 seconds on a desktop.
Can I keep a shadow, or cut out something that is not the main subject?
No. The model picks the salient subject and everything else becomes transparent, with no brush, no eraser and no way to tell it which object you meant. For that degree of control you need an editor with manual selection.
Does it work offline?
After the first run, yes. Once the model is in your browser cache the tool needs nothing from the network, so the page keeps working with the connection switched off.
Related tools
Other tools people use alongside the Remove Background.
IMG-008Resize ImageChange image dimensions online, with or without locking the aspect ratio.Open
IMG-007Compress ImageReduce image file size online without leaving your browser.Open
IMG-003PNG to WebP ConverterConvert PNG images to WebP online for free.Open
Browse all image tools or see every tool.