Background remover

How to remove a background

1. Add your picture. Drop it anywhere on the page, paste it, or click the box to choose a file. JPEG, PNG and WebP always work.
2. Wait about ten seconds. The first time, the page downloads a model and your browser keeps it, so later pictures skip the download. The status line shows the progress.
3. Check the result. A checkerboard shows the parts that are now transparent. Hold Hold to see the original to compare.
4. Adjust if you like. Pick a different background, soften or sharpen the edge, or switch the picture type.
5. Download. The file is named after your picture with -no-background added, and it keeps the picture's full size.

Photo or anime: two models

The tool uses a segmentation model, a neural network that decides for every pixel whether it belongs to the main subject. There are two of them, because photographs and drawings are different problems.

Auto, the default, picks for you. If one exact color covers 40% or more of the picture, as with a drawing on a plain background, it uses the anime model. Photos almost never do that, because noise and light gradients spread their colors out. The note under the preview says which model was chosen and how big its download is. Choose Photo or Anime or drawing yourself if Auto guessed wrong, and the tool runs the other model. Results are kept for the picture, so switching back is instant.

Both models are pinned to one specific version, and both are stored in your browser after the first download, in the same cache the taggers use. We chose them partly for their licenses: some popular background removal models are free for personal use only, and these are not.

Background, edge and file type

• Background: Transparent, White, Black, or any color you pick. The picture is blended into a solid color along its soft edge. A faint fringe of the old background can remain on fine detail, which the Edge slider helps with.
• Edge: the slider sets how wide the blend between kept and removed is. Sharper gives a crisp cut, good for logos and flat art. Softer blends the edge, which looks better on hair and fur.
• Save as: PNG keeps transparency and is lossless. WebP keeps transparency and is usually smaller. JPEG cannot be transparent, so it is available only when you have chosen a solid background.

What it does well, and where it struggles

The model finds the main subject by itself. You cannot point at the thing you want, so how well it works depends on how clearly the picture has one subject.

• Works well: a person, animal, product or character against a background that is different from them, and illustrated characters on plain or busy backgrounds. In our own tests, the anime model cleanly separated characters from speed lines and from a plain background, and kept thin details such as a cat's tail.
• Several subjects: if two or three things stand out, the model may keep all of them. A blackboard next to a teacher, for instance, was kept together with her.
• No clear subject: a landscape or a pattern has nothing to cut out, and the result can be almost empty. The page tells you when that happens. Try the other picture type, or a picture with a clearer subject.
• Fine detail: loose strands of hair, fur, glass and smoke are hard for any model, and the edge can come out slightly rough or leave a faint fringe. The Edge slider softens it, and a solid background in a similar color to the original hides it best.
• Low contrast: a subject that looks like its background, such as a white cat on snow, is harder to separate.

Speed, size and memory

On our test computer, running the model on plain WebAssembly took roughly ten seconds per picture for either model, once the model was downloaded. Phones and older computers take longer. Pictures are worked on at up to 4096 pixels on the long side, and a larger picture is scaled down to that before it is cut and saved. The model itself looks at a 1024 × 1024 version of your picture, and the mask is then enlarged to the full size, so very fine edges on a large photo are as sharp as that mask allows.

The tool handles one picture at a time. Press Another picture to start the next one.

Uses

• Stickers, avatars and product pictures on a transparent or white background.
• Reference art: isolate a character to paste into another picture.
• Training sets: some people put the subject on a plain background so that a model does not learn the scenery, while others keep varied backgrounds on purpose. See the dataset workflow guide for how that fits with tagging and cleaning, and tag the result afterward with the batch tagger.
• Privacy first: a cutout also strips the photo's metadata, because the saved file is a new image. Our EXIF guide explains what that removes.

Privacy

The model runs inside this tab, and your original file on disk is never changed or deleted. The page downloads a model from Hugging Face the first time you use it. For what the site collects, see our privacy policy.