Remove an Object from a Photo with AI

How it works

You supply two things: the photograph, and a mask marking what should go. The mask is the instruction, and it does more to determine the result than any setting on the page.

The masked region is handed to Qwen Image Edit, an image editing model, together with the surrounding photograph as context. It repaints the marked area from scratch, guided by what is around it, at a denoise of 0.85 over eight sampling steps. A denoise that high means the masked pixels are effectively replaced rather than adjusted — the tool is not cleaning up what was there, it is painting something new in its place.

Everything outside the mask is left as it was.

The mask is the whole job

Two habits account for most of the difference between a clean removal and an obvious one.

Mask generously. Include the object's shadow, its reflection, and a margin of the surface around it. An object removed while its shadow stays is worse than the original, because a shadow with nothing casting it is a thing people notice instantly even when they cannot say why. The same goes for a reflection in a floor, a table, or water.

Mask in one piece where the object is one thing. A mask split into fragments asks the model to solve several small problems independently, and the seams between those solutions rarely line up.

What the surroundings have to provide

The model rebuilds the masked area from context, so the question that predicts success is whether the answer is visible elsewhere in the frame.

Removing a bin from a stretch of pavement works well: the pavement continues on all sides, and there is nothing to invent beyond more of it. Removing a person from a plain wall, an empty sky, sand, grass, or still water works for the same reason.

It gets harder when the object hides something structured. A figure standing in front of a window frame, a doorway, a bookshelf, or a row of railings means the model has to continue a pattern and align it on both sides. Sometimes it does; often the continuation is slightly off, and a line that does not meet is very visible.

It fails when the object hides something unique. Whatever was behind a person standing in front of a face, a sign, or a specific object was never recorded, and what appears there is invention with no evidence behind it.

Removing a watermark is not what this is for

A watermark is placed on an image to mark ownership, and removing one does not transfer any right to use the picture. The image remains someone else's work, the license terms are unchanged, and stripping the mark makes the resulting use harder to defend rather than easier.

There is also a practical point. Watermarks are usually large, semi-transparent, and spread across the whole frame, which is the worst possible case for inpainting — a huge masked area with the original detail visible underneath it and nothing intact nearby to rebuild from. The results are poor as well as unwise.

Checking the result

Look at the boundary of the mask first, at full resolution. A faint difference in texture, sharpness, or noise between the repainted region and its surroundings is the usual tell, and it becomes obvious in print even when it is subtle on screen.

Then check lighting. The repainted area should share the direction and softness of the light in the rest of the frame; a patch that is slightly too bright or too evenly lit reads as wrong before anyone identifies why.

Then look for the shadow you meant to remove and the reflection you forgot.

Finally, be clear with yourself about what the output is. A photograph with an object removed is no longer a record of the scene, and it should not be presented as one where that matters.