Loading tool…
Paint over a photobomber, a power line or a stray bin and let an inpainting model rebuild what should be behind it. Only the area you select changes; the rest of the photo keeps its original pixels and full resolution.
Processing locally on your device
Overview
Inpainting means filling a hole in an image so that it blends in with its surroundings. Old clone-stamp tools copy nearby pixels by hand. Learned inpainting models instead look at the whole surrounding region and generate new content that continues textures, edges and lighting. This tool uses LaMa (Large Mask inpainting, Apache-2.0), a network built around fast Fourier convolutions. These give it an image-wide receptive field, so it is good at extending repeating structures such as brick walls, fences, grass, sand and sky, even across large masks.
LaMa works on a 512 × 512 pixel input. Rather than shrinking your whole photo to that size, the tool crops a padded square around your selection. The padding is about three quarters of the selection's size on every side, so the model sees enough context. The crop is scaled to 512 × 512 and filled, and the result is scaled back and pasted through a slightly grown, feathered version of your mask. Pixels outside the selection are never touched, so a 24-megapixel photo stays 24 megapixels and keeps its sharpness everywhere you didn't paint.
The model file is the OpenCV Zoo export of LaMa, about 93 MB. It is downloaded once and cached by your browser. It runs on your CPU through WebAssembly, because LaMa's Fourier layers don't run reliably on browser graphics drivers yet, and each removal takes roughly 15–45 seconds depending on your processor. Each removal is a separate pass, so you can clean a photo step by step, undo or redo any step with Ctrl+Z and Ctrl+Y, and compare the current state with the original at any time.
| Scene | Expected result | Tip |
|---|---|---|
| Sky, water, sand, grass | Excellent | One pass is usually enough |
| Walls, floors, fences | Very good | Select the shadow too |
| People in a crowd | Good | Remove one or two at a time |
| Text and logos on flat surfaces | Very good | Paint each line of text generously |
| Objects in front of faces or detailed structures | Fair | Expect invented detail; try several masks |
Drop in a photo (JPG, PNG, WebP or HEIC).
Choose Brush and paint over the object and its shadow, or choose Lasso and draw around it.
Use Erase to trim the selection if you painted too much.
Click Remove selected. The first run downloads the model once.
Repeat for other objects, use Undo if a pass isn't right, and Compare to check against the original.
Download the cleaned photo as PNG, WebP or JPG.
Use cases
Remove tourists walking through the background of a landmark shot. Do it in a few passes, one or two people at a time, for the most natural fill.
Erase cables, bins, cars in the driveway or a stray sign so a property photo looks tidy, while the rest of the image keeps its original sharpness.
Clean up dust, crumbs, price stickers and reflections on a tabletop shot before it goes on a menu or into a shop.
Clear the orange date stamp from scanned film prints, or your own old logo from photos you are reusing.
Pro tips
Answers to common questions about the object remover.
Only the selected area is replaced. The fill is generated at 512 × 512 for the cropped region and scaled to fit, so a very large selection in a very large photo can look slightly softer than its surroundings. Everything outside the selection keeps its original pixels.
The first removal downloads the LaMa model (about 93 MB). Your browser caches it, so later removals start straight away and also work offline.
Roughly 15–45 seconds per pass, depending on your processor. LaMa is a large network and runs on the CPU through WebAssembly, which works on every device but is slower than a graphics card would be.
Yes. Undo and Redo step through up to 12 previous versions, and Ctrl+Z, Ctrl+Shift+Z or Ctrl+Y work too. Compare shows the original next to the current result.
Yes, if the background behind them is reasonably simple, such as grass, wall, sand or sky. Busy scenes with people in front of detailed buildings are harder; several passes and a generous mask help.
No. ToolEnhancer processes files on your own device inside the browser. They are never sent to, stored on or seen by our servers.
Yes. The object remover is free to use with no sign-up, no watermarks and no daily limits.