Make an image background transparent (AI cutout)
Automatically remove the background from photos of people, products or pets and save a transparent PNG. The image never leaves your device.
Why these settings
The AI separates the subject in the foreground from everything else and makes the rest transparent. The model is IS-Net, an image segmentation network that handles more than people — products on a table, pets, plants, anything with a clear outline. Unlike colour-based removal, the background does not need to be plain.
The first time, the model (about 91 MB) is downloaded. After that it is kept on the device, so you only wait for the detection itself: about a second per image on a computer with a usable GPU, ten seconds to a minute without one. Either way the image is not sent anywhere.
After the cutout you can still rotate, crop and adjust brightness in the same editor. The background is removed once, first, and later adjustments are applied on top while keeping the transparency. Export as PNG or WebP — JPEG cannot store transparency, so it is disabled here.
Photos without a clear subject are hard: crowds, a subject the same colour as its background, glass or smoke you can see through. The boundary of what stays wobbles. Fine detail such as the ends of hair is kept as partial transparency. If the result is not right, "Restore background" undoes it.
Common questions
- Is my photo sent anywhere?
- No. The AI model is loaded into the browser and the detection runs on this device. The only network traffic is the first download of the model file itself, not your photo. The cut-out image is only written to your device when you press save.
- Why can I not save the cutout as JPEG?
- JPEG has no way to record transparency. Saving as JPEG would fill the removed area with white or black and undo the point of cutting it out. Use PNG to keep it transparent, or WebP if you also want it small — both keep transparency.
- Can I cut out several images?
- Add as many images as you like and press "Remove background with AI" on each. The model loads only once, so from the second image you only wait for the detection. Each image remembers its result, so switching between them does not redo the work.