Is this background remover free?
Yes. This is a free AI background remover with no account, no watermark, and no upload queue. You export a transparent PNG locally.
Does my image stay on my device?
Yes. Upload, the AI subject cutout, preview, and export all happen in your browser, so the image is never sent to a server.
What format does the output use?
The background remover exports a transparent PNG, so the cutout subject drops cleanly onto any background you use it on.
Will every image cut out perfectly?
No browser-side cutout is perfect on every source. Clear subjects, simpler backgrounds, and clean edges usually give the strongest result.
What should I use after removing the background?
Usually the next step is replacing the image background or cleaning up edges if you need more polish or a different output.
Is this really AI, or is it deleting one background colour?
It is a matting model. The tool loads MODNet through transformers.js and runs it in a Web Worker, and the model returns an alpha value for every pixel. Nothing is keyed by colour, which is why there is no tolerance setting. It also explains the failure mode: a dark jacket against a dark doorway is hard because the model has little contrast to work with, not because a colour rule missed.
Does the image get uploaded, and does the tool work offline?
Your image never leaves the device. It is read into a canvas, passed to a local worker, and exported from the same browser tab, with no server round trip. The model weights themselves are fetched once from a public CDN on the first run and then kept in the browser cache, so the very first cutout needs a connection and later ones reuse the cached copy. The picture is never part of that request.
Why is the first cutout slow and the next one fast?
The first run has to download and initialise the model. After that the pipeline stays in memory for the tab, so the second image skips straight to inference. Speed also depends on acceleration: if your browser exposes WebGPU the tool uses it, and if not it falls back to an 8-bit quantised WASM build that runs on the CPU and is noticeably slower. The status line after each run names which one was used.
Why did the gap between an arm and the body get filled in?
A cleanup pass fills fully enclosed transparent regions that are smaller than roughly 0.014% of the picture and do not touch the border. It exists to close pinholes in the mask, but it also closes genuine negative space: small gaps between fingers, the space inside a handle, or a narrow gap under an arm. Crop tighter before you upload so that the gap is a larger share of the frame and survives the pass.
Why does hair come out chunky?
The predicted matte is hardened before export: alpha at or below 28 is snapped to fully transparent, alpha at or above 224 is snapped to fully opaque, and isolated foreground fragments smaller than about 0.018% of the image are deleted. Flyaway hair strands are exactly those fragments. Hair on a high-contrast, evenly lit backdrop survives; hair against a busy or similarly toned background does not, and no browser-side model fixes that.
What kind of background works best?
Contrast matters more than simplicity. The model needs the subject to differ from what is behind it in brightness or colour. A person on a busy but bright street often cuts out better than a person in a dark suit against a dark wall. If a result is poor, changing the background contrast in the source beats retrying the same file.