Remove People 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 people and rebuild the ground, background, and architecture behind them so the place reads as empty and continuous, preserving the location, the perspective, the lighting, the shadows of remaining objects, and the composition, and adding no new figures.

The negative instruction names this job's specific failures — remaining figures, ghosted silhouettes, orphaned shadows, warped architecture, blurred patch.

The repaint runs at seventy five per cent strength, the highest of the removal workflows here, because a figure occludes far more than a small object does and the area behind it has to be genuinely built. The seed is not fixed, so every run produces a different reconstruction.

A person leaves four things

The outline is only the first of them, and it is the one the model handles best.

The shadow is second. A figure standing in sunlight lays a long dark shape across the ground, often several times their own area, and it does not resemble a person at all — it is a stretched gray patch on paving. Removed badly, it stays where it was, and the eye reads a shadow with no owner immediately even when it cannot say what is wrong.

The reflection is third. Shop windows, glass doors, wet pavement, still water, polished floors, vehicle panels. A figure removed from the pavement but still standing in the window behind it is a strange picture.

The occlusion is fourth and the most expensive. Whatever they were standing in front of is missing from your file entirely: the middle of a bench, three columns of a facade, half a doorway. That has to be constructed, and unlike grass or sky it has structure that must line up on both sides.

Checking a result means checking all four, and only the first is likely to be right on its own.

Coverage decides the outcome

There is a straightforward relationship worth knowing before you upload.

The fraction of the frame the people occupy is roughly the fraction of the result that is invented. Two walkers at the far end of a beach cover perhaps one per cent of the image, so ninety nine per cent of what comes back is a reconstruction of your photograph with a small local repair.

A crowd across the foreground of a plaza can cover a third of it. Now a third of the returned image is composed rather than observed, and it is composed to be plausible for a plaza rather than faithful to yours. Steps appear at a slightly different spacing. A doorway lands where a doorway would sensibly land.

Neither result is a failure of the tool. They are the same operation applied to very different amounts of missing information, and knowing which one you are asking for sets what you should expect.

The photograph changes meaning

Worth saying, because it is the reason some emptied images never look right no matter how clean the reconstruction is.

A photograph containing people is usually composed around them. The framing gives them room, the exposure is set for them, the moment was chosen because of what they were doing. Take them out and the composition is left pointing at nothing, with an oddly empty middle and a subject that has gone.

An empty street can be a fine photograph. It is a different photograph, and it is usually shot deliberately rather than made by subtraction.

If the result feels wrong while every individual region looks correct, this is generally why, and it is a framing problem rather than a rendering one.

Publication gate

This page ships once the workflow has been run against distant figures on sand, a crowd in front of a structured facade, a subject casting a long shadow, and a scene with figures reflected in glass, with each output checked for shadows and reflections left behind.