Remove objects from photos

How to remove an object from a photo

1. Add a picture. Drop one anywhere on the page, paste it, or click the box to choose a file. Phone photos, including iPhone HEIC, work too.
2. Paint over what you want gone. Use the brush to cover the object, a person, a sign, a blemish or a wire. A red tint shows what you have marked. Erase a mark takes paint back, Undo stroke removes the last stroke, and Clear marks starts again.
3. Press Remove what I painted. The first time, the page downloads the model (27 MB for Fast, 208 MB for Best quality), and your browser keeps it.
4. Look at the result. Switch the view to Original to compare. If you do not like it, Undo last removal brings the picture back as it was, and you can try again with a different brush, a different quality or a smaller mark.
5. Remove more, or save. You can keep painting and removing as many times as you like, and each removal builds on the last. Then press Download the picture.

Painting well

• Paint a little beyond the object. Include its edge, its shadow and any reflection. If you paint exactly along the edge, a thin rim of the object is left in the picture, and the model treats that rim as part of the scene and tries to continue it, which can give odd scribbles. Cover the edges (from 0 to 20, starting at 4) grows your painted area by a few pixels to catch this. In our test, a stroke painted about 30 pixels too narrow for a disc left 15,597 pixels of its red rim behind with the setting at 0, and none with it at 20.
• Brush size is in screen pixels, so the brush looks the same whatever the zoom. Zoom in with + to paint a small object precisely, and use Fit to see the whole picture. Scroll to move around a zoomed picture. Ctrl + Z (or Cmd + Z) undoes the last stroke.
• Several separate objects are fine in one go. Each is filled in its own area, one after the other, and the page tells you how many areas it filled.
• Smaller is better. The model fills a window around your marks at a fixed size of 512 × 512 pixels. The smaller the marked area is compared with the picture, the sharper the fill. Remove a large thing in a few steps rather than all at once.

Fast or Best quality?

We compared them on pictures where we knew exactly what should be under the marked area, by measuring the average difference from the true picture in the filled area (a lower number is closer, out of 255):

So on smooth or loosely patterned backgrounds, such as sky, water, walls, grass or a blurred background, Fast is just as good and much quicker. On regular patterns such as fences, bricks, tiles, stripes and railings, Best quality is much closer, because it continues the pattern better. Try Fast first. If the fill looks wrong, undo and try Best quality.

What stays the same

Only the area around your marks can change. We checked this on a 1200 × 900 test picture: after removing a disc, all 1,022,791 pixels further than 135 pixels from it were byte for byte identical to the original (the largest difference was 0), and on a 4000 × 3000 picture the 11.5 million pixels beyond 400 pixels from the object were also unchanged. The filled area blends into its surroundings over a few pixels. The picture is saved at its full size, so a large photo keeps all its resolution everywhere you did not touch.

How well does it work?

• Removing a red disc from a smooth, structured background: the average difference from the true background fell from 123 to 6.7 levels with Fast, and to 6.9 with Best quality, in about 2 seconds and about 23 seconds.
• A small object in a 12-megapixel picture: from 122 to 4.1, in about 4 seconds, with the rest of the picture untouched.
• The page gives exactly what the models give. We ran the same marked area through each model natively on a computer and through the page, and the filled pixels were identical, for both models.

What it cannot do well

• It invents what is under the mark. The fill is the model's best guess of what could be there, not what really was there. It is convincing on backgrounds, and unreliable on detail it cannot know: a face, text, a particular pattern, or an object that was half hidden.
• Large areas look softer. A window is filled at 512 × 512 pixels and then enlarged to fit your picture, so the fill in a large area has less fine detail than the photo around it.
• Repeating patterns and straight lines can come out slightly bent or broken, especially with Fast.
• Very large selections (most of the picture) do not work: there is not enough left for the model to learn from.
• It is for your own pictures. Do not use it to take another person's watermark or credit off their work, or to change a picture that is meant as a record, such as a document or evidence.

Saving and limits

• JPG, PNG or WebP. A JPEG picture is saved as a JPG by default, and anything else as a PNG. The file is named after the picture with -removed added. A picture with see-through areas gets a white background.
• Undo. The last three removals can be undone, one by one.
• Very large pictures. A picture over 24 megapixels is made smaller when you open it, and the page says so.
• One picture at a time. Your original file is never changed.

Ideas for using it

• Clean up a travel photo: a stranger in the background, a sign, a trash can or a power line.
• Retouch small flaws: a spot, dust on a scan, a stray hair or a date stamp on your own picture.
• Tidy a product photo before putting it on a plain background with the object cutter.
• Prepare a training picture: remove a logo or text from your own image before it goes into a dataset.

How it works

The page marks the pixels you painted, grows them a little, and takes a square window of the picture around them, with plenty of the surroundings. The window is scaled to 512 × 512 pixels, and a model fills in the marked pixels so that they match what is around them. The result is scaled back to the window's size and blended into your picture with a soft edge. Both models, MI-GAN from Picsart and LaMa from Samsung's AI research center, learned from a large collection of photographs of places.

Privacy

• The removal happens inside this tab, and your files on disk are never changed or deleted.
• The model files are downloaded the first time from Hugging Face, a model hosting site, and your browser keeps a copy.
• For what this site collects and keeps, see our privacy policy.