Remove Unwanted Objects from a Photo

How it works

Your photo is scaled so its total area is around one megapixel, keeping the original proportions. An image-editing model reads that copy along with a fixed instruction: remove the unwanted objects and reconstruct the surfaces behind them so the scene reads as continuous, preserving the people, the remaining objects, the perspective, the lighting, and the composition, and adding nothing new.

A second instruction names the failures this job actually has — altered subject, changed lighting, warped perspective, blurred patch, cloned artifacts, broken pattern, extra objects, text, watermarks.

The repaint runs at seventy per cent strength, which is higher than the restoration passes on this site because there is real content to build rather than damage to clean. The seed is not fixed, so each run invents a different filling.

There is no ground truth

This is the thing to hold on to when judging a result.

A restoration tool has a target. The scratch covered something that existed, the noise obscured a texture that was really there, and success means getting closer to a thing the photograph is evidence of.

Removal has no such target. The wall behind the bin was not photographed. No information about it exists anywhere in your file. So the model does not recover it — it composes something that would be consistent with everything around it, using what it knows about walls.

That reframes the question. Asking whether the fill is accurate is meaningless, because there is nothing for it to be accurate to. The only question available is whether the fill is consistent, and consistency is something you can actually check.

What to check, in order

Four things, and they take a few seconds.

Repeating patterns first. Tiles, bricks, floorboards, fence slats, panelling, railings. These have a rhythm your eye can count along, so a fill that changes the spacing or drops a unit is visible without effort. This is the single most common giveaway.

Straight lines second. A skirting board, a table edge, a window frame, the join between wall and floor. A line that enters the filled region and leaves it at a slightly different angle is a failure that people notice without knowing why.

Perspective third. Anything receding towards a vanishing point must keep receding at the same rate. The instruction names warped perspective for a reason.

Shadows last. The removed object may have been casting one, and a shadow left behind by something that is no longer there is unmistakable once seen.

Easy backgrounds and expensive ones

The difficulty of a removal is almost entirely a property of what was behind it.

Grass, foliage, gravel, sand, water, sky, wood grain, concrete, and carpet are close to free. They are irregular by nature, so an invented patch has nothing to be checked against — there is no correct arrangement of blades of grass.

Tiles, brickwork, text, patterned fabric, bookshelves, railings, and faces are expensive. Each carries structure a viewer can verify, and a wrong reconstruction is not merely imperfect, it is legible as wrong.

The practical version: before running this, look at what is behind the thing you want gone and decide whether an invented version of it could be checked. If it could, expect to run several attempts, and expect some of them to be unusable.

Publication gate

This page ships once the workflow has been run against an object on an irregular surface, an object against a tiled wall, an object casting a visible shadow, and a scene with strong perspective lines, with each fill checked for pattern continuity and shadow consistency.