Remove Background from Images

How the AI decides what to keep

Background removal is a job for a segmentation model: a neural network trained on many thousands of photos to estimate, for every pixel, how likely it is to belong to the main subject. The result is a soft mask, fully opaque where it’s sure, fully transparent where it’s sure it’s background, and partly see-through along edges like hair and fur.

ImageCocoon uses two small open-source models and picks between them for you. Photos with a clearly visible face go to MODNet, which was trained specifically on people and handles curls, flyaway hair and shoulders well. Everything else, such as products, pets, plants and objects, goes to U²-Net, a general-purpose model. A tiny face detector makes the choice in a fraction of a second. If it picks wrongly, set “What’s in your photos?” to People or Products yourself.

Transparent, colored or blurred

• Transparent saves a PNG (or WebP) with a see-through background, ready to drop into slides, designs, documents and stickers. See Make Background Transparent for tips.
• Color places the subject on any solid color you pick. White is the standard for online shops and many ID-style photos; there’s a ready-made white background page.
• Blurred keeps the original scene but softens it behind your subject, like a phone’s portrait mode. Choose Even for the same blur everywhere, or Lens, where a depth AI blurs things more the farther away they are. Try Blur Background.

Changing the background, color or blur after the first run is quick, because the AI’s mask is remembered for each photo; only the final picture is rebuilt.

Getting clean edges

Hair, fur and fuzzy edges are the hardest part of any cut-out. The Edge cleanup slider decides how strict the tool is with partly transparent pixels. Raising it removes faint halos of the old background, which is useful when someone stood in front of a bright wall. Lowering it keeps more wispy detail, which looks better on a similar-colored new background.

For anything the AI got wrong, the touch-up brush is quicker than starting over. Choose Erase to paint away leftover background, or Restore to bring back an ear, a shoe heel or a product edge that was cut off. Turn on Tap a part and a single tap erases or restores a whole piece at once, such as a chair behind someone or an arm the AI dropped; a separate object selector finds its outline. Adjust the brush size, undo mistakes step by step, or reset to the AI’s original mask. Touch-ups are kept per photo, so you can fix one image in a batch without affecting the others.

What works well, and what doesn’t

Expect excellent results with portraits, pets, products on a table or floor, cars and most single, clearly defined subjects. Results are weaker when the subject and background share colors, with glass and transparent objects, with very thin parts like bicycle spokes, and with busy scenes containing several things that could each be “the subject”. Group photos usually keep everyone, but sometimes also keep the ground they’re standing on.

Our models are compact so they can run on ordinary laptops and phones. Large commercial services use much bigger models on powerful servers and can be more precise on difficult hair, but they require uploading your photos. We think a few seconds of touch-up is a fair trade for keeping your pictures private.

Privacy and the AI download

Nothing about your photos is sent anywhere. What gets downloaded, once, is the AI itself: the runtime that executes the models (about 14 MB) and the models (about 11 MB), all served from our own domain and checked against a fingerprint before use. Your browser stores them, so the next visit starts instantly and works even with a weak connection. If you ever want to remove them, clear this site’s data in your browser settings. For more on what to look for in online tools, read whether online image tools are safe.