Object Remover (Magic Eraser)

About the Object Remover (Magic Eraser)

Remove a stranger who walked into your shot, a bin, a sign, a power line, a pimple or a date stamp from your own photo. Paint over it with the brush and select Remove brushed area: an AI model fills the space with what plausibly belongs there — sky, grass, wall, skin or water. Brush over anything that is left and remove again, compare with the original at any time, and undo any step.

The model is MI-GAN, a fast inpainting network made for phones, and it runs in your browser: the photo is never uploaded. The area round each removal is filled at the photo’s own resolution when it is up to about 256 pixels across, so the rest of the picture keeps every pixel of its original quality.

Common uses

• Removing photobombers and passers-by from holiday photos.
• Cleaning up wires, signs, litter and parked cars in landscape and property photos you took.
• Retouching small blemishes and stray hairs in portraits.
• Removing date stamps and dust spots from your own scanned photos.

Getting good results

• Cover the whole object, edges and shadow included: anything left outside the brush stays in the photo, and the model copies traces of it.
• Small to medium areas work best. For a big object, remove it in a few passes — a part at a time — rather than all at once.
• Plain or natural backgrounds (sky, grass, sand, walls, water) fill most convincingly. Regular patterns, text, straight lines and other people’s faces are harder: expect to touch up.
• If a fill looks wrong, Undo, brush a slightly different shape and try again: each pass can give a different result.

How it works

MI-GAN (Sargsyan et al., ICCV 2023, by Picsart AI Research; MIT licence for the code and the trained weights) predicts the missing part of a picture from what surrounds it. MySmartCoPilot sends it a square of the photo round your brush strokes — at least 512 × 512 px (or the whole photo, if it is smaller), with 128 px of context on each side — and writes the result back only where you brushed, slightly softened at the edge, so every pixel elsewhere stays exactly as it was. The model works at 512 × 512 internally: areas larger than about 256 px across are filled at a lower resolution and look softer than the rest of the photo.

Use it responsibly

Edit photos that are yours or that you have permission to change. Removing watermarks, logos, signatures or photo credits from someone else’s image can infringe copyright and break the image’s licence — and many stock and news photos forbid it. Do not use edited photos to mislead, for example in documents, evidence, news or listings that must show things as they are.

Limitations

• The fill is a plausible guess, not the real hidden background; complex structures and faces may need several tries.
• Areas larger than about 256 px across are filled at a lower resolution than the rest of the photo.
• Photos larger than 16 megapixels (8 on phones and tablets) are scaled down to that size for editing.
• The first removal downloads about 42 MB; later removals use the cached copy.
• Edits are kept only while the page is open.

Privacy

Your photo is processed in your browser and never sent to a server. The first removal downloads the AI engine and the object-removal model (about 42 MB) from MySmartCoPilot, not from a third party, and your browser’s cache keeps them for next time. The saved image is encoded fresh, so it carries no EXIF data from the original (no camera details or GPS location).