AI Powered · Free · No Signup

Remove Background Free

Get a clean transparent PNG in a few seconds. One model is trained on people, another on everything else, and the tool picks for you. No signup, no watermark, no limits — the image never leaves your device.

Drop an image to remove its backgroundor click to browse · JPG, PNG, WebP · AI powered
Auto tries the people model first and switches if the photo is not a portrait.
0% keeps soft AI edges. Increase for a sharper cut — may lose fine details like hair.
0%
Softens rough edges. Try 1–2px for natural results.
0px
Negative trims edges tighter. Positive keeps more around the subject.
0px
Transparent for design tools. Solid colour for documents.
#ff7f3d
PNG for transparency. WebP for smaller files.

Click the subject you want to keep.

Click the subject to keep

Each click is instant — the slow part already ran. Add clicks to extend the selection, or shift-click (or right-click) to cut a part back out.

Original Original
Background removed Background removed

The result is shown on a checkerboard so you can see which areas are actually transparent. Adjusting the sliders re-composites instantly — the AI does not run again.

Freeno limits, no watermarks
AIneural network powered
100%private — runs in browser

Two models, because one does not fit everything

Portrait matting and general object segmentation are different problems. A model trained on people learns hair, skin edges and clothing, and is small and quick because that is all it has to know. A general model has to handle a shoe, a plant and a car, so it is bigger and slower but does not fall apart on things that are not human.

QuickImg ships both and chooses for you. The people model runs first because it is a fraction of the size; when it comes back with an essentially empty cut-out — which is what it does on a photo with no person in it — the general model takes over. You can also force either one if the guess is wrong.

Getting a clean cut

Frequently Asked Questions

Which AI models does this use?▸
Two, both Apache-2.0: MODNet for people, and Meta’s Segment Anything (SAM, in its compact SlimSAM build) for products and other objects. MODNet finds a person by itself; SAM segments whatever you point at, so object mode asks you to click the subject. Both run in your browser through WebAssembly, or on your GPU through WebGPU when the browser supports it.
How does it know whether my photo has a person in it?▸
It runs the people model first, which is small and quick. That model is trained only on portraits, so on a photo without a person it returns a nearly empty cut-out. When that happens the tool loads the general model and runs again. You can override the choice with the Subject setting.
Is the image really not uploaded anywhere?▸
Correct. The model files are downloaded from a CDN to your browser, then all inference happens locally. Your image data never leaves the device — you can confirm it in your browser devtools Network tab, where no request carries your file.
Why is the first run slower?▸
The first time, the browser downloads the model: about 6 MB for the people model, about 13 MB for the general one. They are cached afterwards, so later images skip straight to processing. Processing itself runs in a background thread, so the page stays responsive.
Can I adjust the result?▸
Yes. After the cut-out is generated you get threshold, smoothing and shrink/grow controls, plus background colour and output format. These re-composite from the stored mask instantly, without re-running the AI.

Other Tools