Remove Person From Photo: What Actually Works

Nothing gets deleted. The pixels a person occupied have to be replaced by something, and that something is manufactured. Every remove-person-from-photo tool is doing inpainting (2026): reading the surviving surroundings and writing a continuation that looks like it belongs.

Two families of tool do this. Patch-based erasers copy texture from elsewhere in the same image, which works beautifully on grass and terribly on a face-height gap in a shop window. Generative models write new content from scratch, which handles complicated backgrounds and occasionally hallucinates a doorway that was never there.

Set the standard now, before any of the how-to below. A removal is judged at 100 percent zoom on the join, not at thumbnail size in a feed. Almost everything passes at thumbnail size, which is why so many obviously edited photos circulate.

The workbench route has no lasso and no brush. You describe what leaves in the Prompt box, which is unfamiliar at first and turns out to be more precise than a wobbly finger-drawn selection, because you can also state what has to survive.

1. 1

   Load the photo

   Open the object remover and drop the file into the Photo box. Use the original, not a messaging-app copy, because compression artefacts get rebuilt along with everything else.

2. 2

   Describe the person by position, not by name

   "The woman walking on the left pavement" is findable. "My colleague" is not. Position, clothing colour and direction of travel are the three attributes that land reliably.

3. 3

   Include the shadow and any reflection

   This is the step people skip and the reason removals look wrong. A body leaves a shadow on the pavement and sometimes a smear in a shop window. Name both or they stay behind as evidence.

4. 4

   State what must not move

   Add a clause protecting the parts you care about. Without it the pass is free to tidy anything it considers untidy.

5. 5

   Set the ratio, then run

   Keep the original aspect ratio so the frame is not recomposed while you are fixing one corner of it. The price sits on the Generate button.

Remove the woman walking on the left pavement together with her shadow, rebuild the shopfront and paving behind her, leave everything else identical.

That single observation is worth more than any settings advice. "Leave everything else identical" was already in the prompt and the bicycles went anyway. Compare the full frame against the original before you accept a result, and if something else has changed, run again with that object named as protected.

An arm around a shoulder is a different problem from a stranger twelve metres away. The tool has to rebuild the part of your subject that the other person was covering, and there is no reference for it anywhere in the frame.

Three things decide whether this works. How much of the remaining person is hidden: a hand on a shoulder is recoverable, a body half behind another body is not. Whether hair overlaps, because hair edges are where reconstruction is most obviously wrong. And whether the light was directional, since a removed body also removes the shadow it was throwing onto the person you kept.

• **Protect the survivor explicitly.** Name the person staying, their pose and the light, in the same sentence that asks for the removal.
• **Do it in two passes when the overlap is heavy.** First remove, then run again asking only for the rebuilt edge to be cleaned up.
• **Consider the crop instead.** A tighter frame that simply excludes the other person is honest, instant and costs nothing. It is often the better answer and nobody ever suggests it.

The most reliable way to edit someone out of a picture is to crop them out. Everything else is reconstruction.

— Maya Chen

This is the easy end of the job and the reason the whole category exists. A tourist against open sky is trivial: the replacement is more sky. Difficulty scales with how structured the background is.

For repeating architecture, give the tool the rule rather than the outcome. Telling it to continue the railing at the same spacing across the gap produces better joins than asking it to remove the man in the blue jacket, because the second instruction says nothing about what should replace him.

It is the most-searched version of this job and the one with the most ways to go wrong, because the giveaways are not in the gap. They are in the person who stayed.

• **Colour spill.** A red shirt standing next to someone throws warm light onto their cheek and shoulder. Remove the shirt and the tint remains, pointing at a person who is no longer there.
• **A contact shadow.** Where two people touch, there is a dark line. Removing one leaves a shadow with nothing casting it.
• **Grain mismatch.** The rebuilt region is usually cleaner than the rest of the photo. On a phone shot taken indoors, that patch of unnaturally smooth wall is the tell.
• **Pose logic.** An arm that was around a waist ends up floating. Ask for the arm to be repositioned naturally, or crop above it.

Practically, run the removal, then run a second pass that asks only for the lighting and grain around the join to match the rest of the frame. Two 13-credit passes on a photo that matters is a reasonable trade, and the second one is where the result stops looking edited.

Every image tool in the workbench runs on GPT Image 2, so the same instruction style carries across removal, background work and repair. That is useful here: the vocabulary you learn writing removal prompts transfers directly.

There is no single remove people from photos app that wins on both platforms, because the best option on each one is the one already installed.

The tap-to-erase apps are quicker and the prompt route is more controllable, and the difference shows up exactly where you would expect. A tap knows where you pointed. It does not know that the shadow two feet to the left belongs to the same person.

Several, and they differ in what they take away rather than in what they charge. Free remove-person-from-photo tools generally clip the output size, add a mark, or cap how many you can run. Everything quoted here was taken from the tool's own site on 7 September 2026.

One structural note about the free credits in the image generator: they are not divided per tool. A removal pass, a follow-up cleanup pass and a background swap all come out of the same pool, which is how the two-pass approach above fits inside a new account.

There are categories where a removal stops being a tidy-up and becomes a false statement about a place, an event or an asset.

• **Property listings.** Removing a person is usually harmless, but the same tool removes clutter, damage and neighbours. The NAR Code of Ethics (2026) obliges agents to avoid misrepresenting a property, and an edited photo is a representation.
• **Anything advertising a product or service.** The FTC's endorsement guidance (2026) turns on whether a claim is misleading, and a photograph makes claims.
• **News, documentary and insurance.** A removed bystander is a removed fact. If the image is describing what happened, do not change who was there.
• **Legal and identification material.** Passport, visa and court exhibits have their own rules, and generative editing breaks all of them.
• **Photos of other people you did not ask.** Removing someone from a shared memory is a social act as much as a technical one.

For everything else, which is most things, removal is fine and nobody needs a disclaimer to enjoy a holiday photo without a stranger in it. The line is simple: if the picture is a record, leave it alone. If it is a picture, edit it.