Remove background from image

Where a cut-out image is useful

Removing the background turns a photo into a layer you can place anywhere. Typical uses:

• Product photos for a shop or marketplace that requires a plain white or transparent background.
• Profile pictures and team pages where everyone should share the same backdrop.
• Thumbnails, posters and slides where a person or object sits on top of a design.
• Stickers and cut-outs for messengers or print.

The result is a PNG in which the subject keeps its original pixels and everything around it becomes transparent, so you can put it on any colour, gradient or photo afterwards.

How the subject is found

The tool uses ISNet, a segmentation network trained to separate the main foreground object from its surroundings. It runs through the ONNX Runtime engine compiled to WebAssembly, entirely inside the browser tab. The model file is a compact 8-bit version of about 45 MB; it is downloaded on the first run and kept in the browser cache, so later visits start processing straight away.

For each image, the process is:

1. The photo is scaled to 1024 by 1024 pixels, the size the network was trained on.
2. The network produces a mask where each pixel gets a value between fully background and fully foreground.
3. The mask is scaled back up to your image's original dimensions with smooth interpolation.
4. The mask becomes the alpha channel of the original photo, and the result is saved as PNG.

Two consequences are worth knowing. First, the output keeps your original resolution, because only the mask is computed at 1024 pixels; the subject itself is never downscaled. Second, the mask is soft rather than strictly on or off, so edges such as hair, fur and motion blur fade out gradually instead of being cut with a hard line. That usually looks more natural, but it also means semi-transparent fringes can carry a trace of the old background colour.

Photos that work well and photos that do not

The network looks for one clear subject. You get the cleanest results when:

• The subject is in focus and takes up a good part of the frame.
• There is contrast between subject and background in colour or brightness.
• The subject is a single person, animal, product or vehicle.

Harder cases:

• Busy scenes with several objects. The model has to guess which one is the subject and may keep parts of the background or drop part of a group.
• Subjects matching the background. A white mug on a white table, or dark hair against a dark wall, gives blurry or missing edges.
• Transparent and reflective objects. Glass, water and chrome show the background through themselves, which no mask can represent cleanly.
• Very wide images. Since the image is squeezed to a square for analysis, a long panorama loses horizontal detail and thin features may be missed. Cropping closer to the subject first with crop image often improves the mask noticeably.

If the result is almost right, a quick touch-up of the edges in any image editor is usually faster than reshooting.

Making the result fit where it goes

Transparent PNG files are larger than JPG, often two to five times the size of the original photo, because PNG compression is lossless and the alpha channel adds data. That is fine for design work but heavy for a web page or a marketplace upload.

• For a solid background, convert the PNG with PNG to JPG, which fills transparent areas with white or any colour you pick and produces a much smaller file.
• For a smaller transparent file, scale it down to the size you will actually display with resize image; a 4000-pixel cut-out shown at 800 pixels is wasted bytes.
• For a consistent product catalogue, shoot all items from the same distance so the cut-outs come out at a similar scale.

You can drop several photos at once. They are processed one after another and can be downloaded individually or together as a ZIP.

Device requirements

Because the model runs on your own processor, speed depends on your hardware. On a recent laptop each image takes a few seconds after the model is loaded; older phones can take considerably longer. Memory use scales with image size: a 12-megapixel photo needs several hundred megabytes during processing, so very large images may fail on phones. If that happens, resize the photo first or use a desktop browser. No image ever leaves the device, which also means the tool keeps working on a slow connection once the model is cached.