Remove Objects from Photos

Two AI models working together

Removing an object takes two steps, and each has its own small model. The first is an object selector based on MobileSAM, a compact version of Meta’s Segment Anything. It reads your photo once, in a couple of seconds, and from then on turns every tap or box into an outline almost instantly. The second is MI-GAN, an inpainting model from Picsart AI Research, which looks at the area around your selection and invents pixels that continue the surroundings: sky, grass, sand, walls, water.

We made the object selector smaller for the web by storing its weights with 8-bit precision, which cut its download from 45 MB to 19 MB without changing its outlines on our test photos. Together with the remover and the engine that runs them, the first visit downloads about 61 MB, once, after asking you.

Tap, box or brush: which to use

• Tap works best on clear, single objects: a bottle on a table, a sign, a parked bike. Tap again to add a part the outline missed, or turn on Subtract and tap to leave something out.
• Box is the most reliable way to select a whole person. A single tap can grab only a jacket or a face; a box around them tells the AI you mean the entire figure. See removing people from photos.
• Brush is for things AI outlines struggle with: thin wires, text, blemishes, shadows and reflections. It is also how you fix a selection by hand. See removing text from an image.

Whatever you select is grown by a few pixels before filling, so the soft edge around an object goes too. Shadows are a separate thing to the AI, so brush over a shadow as well, or it will stay behind as a dark patch.

Getting natural-looking results

Inpainting is guesswork, and it guesses best when the background is simple or repeating. Removing a person from a beach, a field or a plain wall is usually seamless. Removing something in front of detailed, unique content, such as a face behind a lamp post or text over a pattern, is harder, because the model has to invent detail it cannot see.

Work in small steps. Remove one thing, look, then move to the next; several small removals look better than one huge selection. If a fill looks smudged, press Undo, adjust the selection slightly and try again, since each run can come out a little differently. For a very large removal, it can help to crop the photo first with Crop Image so less needs inventing.

Full quality where you didn’t edit

For speed, the editor works on a copy of your photo up to 2048 pixels across. When you save, every removal is replayed on the original full-size image: only the selected areas are replaced, and the rest of the picture keeps its original pixels. The file is saved in the same format as the original (JPG stays JPG, PNG stays PNG), and the image keeps its size.

Private, and fair to use

Your photos are never uploaded. Both models are downloaded from our own domain and run on your device, which also means you can keep working with a weak connection once they have loaded. Please use the tool on photos you own or have permission to edit. Removing other people’s watermarks or credits from their work can break copyright law, and it is not what this tool is for. Read more about keeping photos private when using online tools.