AI object removal feels like it reveals what was hidden behind the object. It does not. Nothing in the photo records what was behind a person standing in front of a wall; those pixels were never captured. The model invents a plausible patch from what it can see around the hole. Everything about when it works and when it fails follows from that.
This guide explains what happens when you press Remove in the object remover, which photos it handles well, which it cannot, and how to get a result nobody notices.
What happens when you press Remove
- You select the object. "Circle it" and "Tap it" use a small segmentation model (SlimSAM, about a 14 MB download) that finds the object's outline from a rough loop or a single tap. "Find objects" uses a detector (YOLOS tiny, about 10 MB) that knows 80 everyday kinds of thing, such as people, cars, bottles, dogs and chairs, and outlines them for you to tap. It misses small or unusual objects; the loop, the tap and the brush always work.
- The selection is grown slightly, by about 0.4% of the photo's longest side, so the soft edge of the object goes too. A leftover halo is the most common giveaway of a removal.
- The shadow is found if "Remove its shadow too" is on, which it is by default: a darker patch touching the object that keeps the colour of the ground around it, since a shadow darkens a surface without changing its colour.
- The model looks at a square around the hole, about three times the size of what you selected, and fills the hole from that context. The square is scaled to the model's working size of 512 pixels when it is larger.
- The patch is fitted back into the photo. Its colours are matched to the pixels around the hole, the photo's own fine grain is added back if the patch was made on a scaled-down copy, and the edge is feathered over a few pixels. Every pixel away from the hole stays exactly as it was, at full resolution.
In Quick mode, thin selections spread over a wide area, such as a power line or a scratch across the frame, are filled piece by piece at full resolution, each piece from its own close surroundings, instead of from one blurred overview of the whole area.
Quick or Best quality
The Fill setting chooses between two models:
- Quick uses MI-GAN, a 28 MB download. A 512-pixel area takes about a second. It is good at random, natural textures.
- Best quality uses LaMa, a 208 MB download that takes 10 to 60 seconds per removal. Its design lets it see the whole square at once, so it continues regular patterns: bricks, tiles, ruled lines, fences, window grids.
The difference was measured on test scenes where the true background was known. On tiles and bricks, LaMa's error was about a third of MI-GAN's. On grass and sky they were the same. So use Quick for nature and plain surfaces, and switch to Best quality the moment there is a pattern or a straight line behind the object.
Where AI fill works well
- Small objects on random texture. Litter on grass, a bird in the sky, a stain on a carpet, a pebble on sand. The model only needs to continue something that has no exact structure.
- Distracting people in the background, when they are small in the frame and standing against something simple.
- Wires, cables and thin poles against sky or a plain wall.
- Spots and marks on skin, clothing or walls, and a date stamp in the corner of your own photo.
- Objects on regular patterns, such as a tiled floor or a brick wall, with Best quality.
Where it fails, and why
- Large objects. When the hole is a big part of its own surroundings, most of the patch is invention with little to anchor it. Expect smears and repeated texture. Removing a person who fills a third of the frame rarely works cleanly.
- Unique structure behind the object. If the object hides half of a face, a sign, a hand or a specific building detail, the model cannot know what was there. It will produce something shaped roughly right and wrong in every detail.
- Objects touching the subject. A bag held against someone's side, a person with an arm around your subject. The fill has to rebuild the edge of the subject, and edges of people are where errors show most.
- Lines in perspective. Floorboards and railings that converge into the distance. Best quality helps, but strong perspective still bends.
- Reflections and detached shadows. Automatic shadow removal only finds a shadow joined to the object. A reflection in a window or a shadow that falls away from the object stays, and it gives the removal away. Select them separately.
- Objects at the edge of the frame. There is context on fewer sides, so the patch has less to go on. Cropping the edge off is often better than filling it.
Habits that make removals invisible
- Select generously. Include the soft edge, any outline and the shadow. A hole that is too tight leaves a ghost of the object around the fill.
- Work in small passes. For a big or complex object, remove one part at a time, starting from the edges. Each pass gives the next one real pixels to work from.
- Use Undo freely. Results vary with the exact selection. If a fill looks wrong, undo and select a slightly different area rather than removing the bad patch again, which would give the model its own mistake as context.
- Compare before and after. Press and hold the compare button to flick between the original and the edit. Your eye catches a repeated texture or a colour step far faster that way.
- Check at full size. Photos up to 24 megapixels are worked at full size on a computer, and up to 12 megapixels on a phone, which has less memory; larger ones are scaled down. Zoom in on the edge of the patch before you save.
- Save as PNG if you will edit further. Each JPEG save loses a little detail. Save as PNG while you are still working, and as JPEG at the end; see image file sizes explained for why.
When not to use it
Official photos must show you as you are. Passport and visa photo rules generally forbid any digital alteration beyond cropping and resizing, and an edited photo can get an application rejected. Do not remove objects from documents such as receipts, certificates or ID scans: an edited document is no longer a true record, and presenting it as one can be a serious offence. For those photos, the rules behind photo uploads are the guide to follow.
The object remover runs entirely in your browser. The models download once and are kept by your browser; your photo is never uploaded.