How do I remove people or unwanted objects from a photo with AI? Prompts that keep the rest intact

Upload the original, say who or what to remove and where, and ask for a natural fill. Success rates from 279 Miku AI users, prompts and a checklist.

Published · 4 min read

Short answer: Upload the original photo as a reference image and write something like "remove the person in the red jacket on the left, fill the area with the surrounding background, keep everything else unchanged". If the area is hard to describe, use the "Mask edit" button on the reference image and paint over what should go. On Miku AI over the last 30 days, GPT Image 2 produced an image in about 86% of 632 remove-object tasks.

How AI removes people from a photo

Classic retouching hides a stranger by cloning nearby pixels over them. Current image models work differently: they read the whole photo and redraw it to your instructions, painting plausible background where the person used to be.

That gives you more convincing pavement, walls or sea, but the model may also touch things you wanted to keep. So a good prompt says what to remove and what must not change.

What Miku AI users actually did

We looked at remove-people and remove-object tasks on Miku AI over the last 30 days (up to November 2026), grouped by prompt keywords:

  • 279 users submitted 1,349 tasks. Every one included a reference image, almost always just the original photo.
  • About 72% produced an image. About 15% were refused by content moderation, mostly photos with real people in them. See why benign requests get refused for how to phrase these.
  • 16:9 was the most common aspect ratio (548 tasks), then 3:4 (325): travel shots and phone portraits.
  • About 60% chose 1K output; 2K and 4K made up the other 40%.
  • 37 tasks used the mask editor to paint the area to remove.

Results by model (only models with at least 20 tasks on this job; time is the median from submission to finished image, successful tasks only):

ModelTasksSuccessMedian timeCredits per 1K image
GPT Image 263286%73 s0.3
GPT Image 2.6 Sunburst4393%49 s0.3
Nano Banana Pro3281%87 s0.35
GPT Image 2.5 Flare17762%60 s0.3
Nano Banana 25756%88 s0.25
GPT Image 2.5 Sunburst36849%74 s0.3

Start with GPT Image 2: it has by far the largest sample and stays above 80%. GPT Image 2.6 Sunburst was faster and succeeded more often, but on only 43 tasks, so treat it as a good second option. GPT Image 2.5 Sunburst did worse on this job, mainly because about 39% of its tasks were refused by moderation.

Step by step in the workspace

  1. Open the workspace and upload the original photo as a reference image.
  2. Pick GPT Image 2 (or GPT Image 2.6 Sunburst).
  3. Set the aspect ratio to match the original, for example 16:9 for a landscape shot or 3:4 for a phone portrait. A different ratio makes the model crop or invent new edges.
  4. Choose 1K for social media, 2K or 4K if you will print it or zoom in.
  5. Write the prompt: what to remove, where it is in the frame, and what to keep.
  6. If there are many people or the position is awkward to describe, click "Mask edit" on the reference image and paint over the people and their shadows in the "Edit mask" editor. Unpainted areas are protected.
  7. Compare the result with the original side by side. Check edges around the people you kept, ground texture, and straight lines such as railings and tiles.

Prompt templates

Text only:

Remove the two tourists walking in the background on the right, along with their shadows, and fill the area with the surrounding stone pavement and buildings. Keep the woman in the white dress in the foreground exactly as she is: face, hair, pose and clothing. Do not change the lighting, colours, composition or aspect ratio.

With a mask:

Remove the people and objects in the painted area and fill it with clean sand and sea that match the surroundings. Keep the horizon level. Leave every unpainted area exactly as it is.

Common mistakes

  • Writing only "remove the people". With several people in the frame the model cannot tell who the subject is, and may remove you too. Say who stays and who goes.
  • Forgetting shadows and reflections. A shadow with no person attached looks wrong.
  • An aspect ratio that differs from the original. You get cropping or a band of newly invented scenery.
  • Removing too much at once. If a crowd covers half the frame, the replacement background is pure guesswork. Two or three passes look more natural.
  • Editing the edited image over and over. Each redraw can lose a little detail; go back to the original when you can.

If you want to change one detail and leave the rest untouched, see how to edit one part of an image and keep everything else.

FAQ

Why was my request to remove people refused?

Photos with real people get more cautious automated moderation, and about 15% of these tasks were refused. Describe the purpose ("clean up the background of a travel photo"), identify people by position and clothing colour, and do not comment on their bodies or looks. Refused tasks are refunded on most models.

Will the filled-in background look fake?

Small strangers, bins and power lines usually blend in well. The larger the hidden area, the more the model has to guess, so zoom in to check and split large removals into several passes.

How many people can I remove at once?

There is no fixed limit, but the more specific you are, the better. For three or more, paint them all with the mask editor, or remove them in batches.

Do failed attempts cost credits?

Credits are charged on submit and refunded automatically if the generation fails. On a few models with a content-policy rule, a failure for a content violation may not be refunded. Reruns are charged normally; see pricing.

More posts