How to Remove an Unwanted Object From a Photo

When a removal is worth doing

The test is simple. Can the background behind the object be inferred from the pixels around it? Open sky, grass, water, asphalt, painted walls and gravel have no fixed shape, so whatever gets rebuilt there usually disappears. If the object hides something specific, half a face, lettering on a sign, a row of railings, the tool has to invent it, and invented detail is the first thing a viewer notices.

The practical cases are the ones everyone runs into. A travel photo you want to keep, with a passer-by standing in the middle of it. A room or a second-hand item you are photographing to sell, with laundry, cables and packaging still on the floor. A slide for a report that needs a background without distractions.

Before you start, make sure the photo is yours or that you have permission to edit it, and save a copy of the original file rather than overwriting it. You need that untouched version to compare against, otherwise you cannot tell whether the repair improved the picture or only smeared it.

The pass I actually make

I work on one object at a time, biggest first, and it usually takes two passes: the body of the object and its shadow go in the first, whatever edge is left over goes in the second. Small nibbling strokes are the fastest way to leave a dark outline behind, so the first mark is deliberately oversized.

1. Save a copy of the original and give the repaired file its own name. Do not overwrite the frame you shot.
2. Zoom in until you can see the object's outline and the shadow it casts on the ground, then set the brush a little wider than the object.
3. Paint the object, its contact shadow and any thin piece sticking out in one pass. The shadow counts as part of the object; it is the mark that gives a removal away.
4. If the object is large, split it. Cover a few hundred pixels at a time, run it, look at the result, then continue with the next piece.
5. Turn the mask overlay off and judge the repair twice: at 100% for the seam and the texture, then at full frame for the tone and the composition.
6. Download the version you want to keep and store it beside the original. If the patch does not hold, go back and repaint the mask instead of painting over the soft area again.

Habits that keep the repair invisible

Failed removals mostly fail at the border, not in the middle. The centre of a patch is usually smooth and convincing; the trouble sits where the rebuilt pixels meet the original ones, and there the shape, the shadow or the grain gives it away.

• Paint the shadow, not only the object. The dark pool under a bag or under someone's feet still reads as the object after the object is gone, and a patch with no shadow looks like it is floating.
• Shrink the brush near straight lines. Kerbs, railings, window frames, floorboards and overhead cables are straight, and people notice when a straight line bends or breaks. Use several narrow strokes rather than one wide one.
• Break large areas into separate runs. Every extra piece you add to the mask gives the tool fewer surrounding pixels to work from. Smaller runs also let you reject one bad piece without redoing the whole area.
• Walk the repeating textures at 100%. Brick, tile, fence wire, fabric and gravel are where duplication shows up, as an element that appears twice on an obvious rhythm or a shadow that suddenly doubles.
• Check the whole frame before you download. Zoom out to the size people will actually see the photo at. A patch that looks fine at 300% can pull the eye at normal size when its brightness or its grain does not match the neighbourhood.

Where it goes wrong, and how to get out

Fine detail that was fully hidden cannot be recovered, only approximated. Leaves and branches, hair, thin metal mesh and printed patterns sitting behind the object come back as a plausible blur, and that blur is obvious at 100%. In low light the repair zone often ends up cleaner than its surroundings, because the surrounding noise is not rebuilt with it. Lettering is a special case: if a word is half covered, the repaired half becomes a meaningless shape, so it is better to remove the whole sign than part of it.

Fix the mask, not the result. If the first run duplicated a railing or grew an extra leaf, go back and paint with a brush narrow enough to stop before that structure, then run it again. If a large sky or wall came back too smooth, run it as two smaller pieces so the join falls where nothing is happening. When a third attempt still looks wrong, change the plan instead: crop slightly to move the object out of the frame, or pick another exposure from the same session, which usually beats a fourth repair.

If the object is covering something a viewer will look for, half a face, a logo or a shop sign, leave it in. A real photo with a stranger in the background usually reads better than a patch that has no explanation.

Clean up first, upscale second

The usual order is removal first, then upscaling. The repair rebuilds texture from pixels that are already in the file, so it should see the sharpest version you have. Run the upscale first and you hand it pixels that an algorithm has already smoothed and, in places, invented; the repair then copies that, and the final image looks softer than it needs to. Removing first also means whatever gets sharpened afterwards is detail you actually kept.

Two situations are worth reversing. If the object is small and sits in a plain area such as sky or a painted wall, cleaning it up is nearly impossible to get wrong at any size, so upscale first and leave the removal as the last step, on the exact pixels you plan to publish. And if you deliver at a fixed large size, enhance a copy just to see whether the defect is even visible at that size; when it disappears at output size, the repair is wasted work.