Removing a background β an AI model running inside your browser
Keep the subject, make the background transparent. The model downloads once; after that the photograph is processed in the browser and never leaves your machine.
Separating a subject from its background means deciding where one ends and the other begins. The old approach keyed on colour and only worked against a plain backdrop. A trained model recognises the outline of people, objects and animals, so a cluttered background is no longer a problem.
The model is fetched once, on first use: 84MB for the standard one, 168MB for the high-quality one. The browser caches it, so there is no wait after that.
The photograph is not uploaded
Most background removal services upload the image, process it on a server and send the result back. Fast and convenient, but the photograph passes through somebody else's computer. For an ID document or a picture of a child, that alone is the problem.
This tool brings the model to the browser and runs it here. The photograph never leaves the machine. Only the model file is fetched from outside, and it carries no information about your image.
There is a second reason the model comes from the library vendor's CDN rather than from this site: serving 84MB to every visitor would be a real bandwidth bill. Not hosting it is part of what keeps the tool free.
Two quality levels
| Standard | High quality | |
|---|---|---|
| Download | 84MB | 168MB |
| Speed | Fast | Roughly twice as long |
| Suits | People, products, clean outlines | Hair, fur, foliage |
| Edges | Good enough | Keeps individual strands |
Why the page does not freeze
Inference is heavy work. Run it in the same place the page is drawn and the interface locks up β buttons do not respond, nothing scrolls, and the visitor concludes the browser has crashed.
This tool moves the computation into a worker, so the page stays alive and the progress bar moves. It uses the GPU where one is available and falls back to the CPU where it is not.
What makes a good result
- β’The more the subject differs from the background in brightness or colour, the cleaner the cut.
- β’Sharp photographs work better; a motion-blurred edge is hard for the model too.
- β’Use the high-quality model for hair and fur β the standard one can clump them.
- β’Translucent things, like a glass or thin fabric, are difficult at any quality.
- β’With overlapping people, someone behind may be treated as background.
Using the result
The output is a PNG with a transparent background: drop it onto another background, put it in a document or a slide, or standardise product photos onto white.
Transparency cannot survive as JPEG. That format has no concept of it, so saving as JPEG fills the transparent area with black or white. Keep it as PNG or WEBP to keep the transparency.
Frequently asked questions
Why is the first run slow?
That is the model downloading β 84MB or 168MB, once. Subsequent runs start immediately.
It stops partway and reports a failure.
Usually the connection dropped while fetching the model. Check the network and try again; the download resumes.
Does it work offline?
Generally yes, once the model has been fetched and while it remains in the browser cache. Clearing the cache means downloading it again.
Can I process several images at once?
One at a time. Each takes a few seconds of computation, and running them together is slower rather than faster.
Is my photograph sent anywhere?
No. Only the model is fetched; the image stays in the browser.