Remove crowds from a travel photo with AI: mask people, keep the place

To remove tourists from a travel photo with AI, paint a mask over the people and send the photo and the mask to openai/gpt-image-2.5 with mask_url, telling it to fill the masked areas with the surrounding pavement or wall and keep everything outside the mask identical. Do the cleanup in zones for a dense crowd.

Mask the people, not the place

Cover each person with a little margin, including shadows and bags. Leave the landmark, signs and the horizon outside the mask. Sume's mask_url is a public HTTPS URL on the GPT Image 2.5 rows, together with up to 16 references. Nano Banana and the other rows do not list it.

The request

Say what to fill with. A vague prompt invites new objects, so name the surface and forbid additions.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-image-2.5","prompt":"Remove the people inside the masked areas. Fill them with the surrounding stone plaza, matching the perspective and lighting. Keep everything outside the mask identical. Do not add people, objects or text.","quality":"high","aspect_ratio":"auto","mask_url":"https://example.com/mask.png","input_references":[{"type":"image_url","image_url":{"url":"https://example.com/plaza.jpg"}}]}'

Work in zones for a dense crowd

A single large mask asks the model to invent a lot of ground at once. Clear the people nearest the camera first, check the result, then mask the next group and run again from that output. Keep the original photo, since each pass is a new generation.

Check what changed

Zoom into the edges of each mask: repeated paving patterns, missing shadows and ghost outlines are the usual faults. Compare unmasked areas with the original to confirm they were not altered. A masked edit is not guaranteed to leave pixels outside the mask untouched, so test it before you rely on that.

Disclose edits where it matters

Removing people changes what the place looked like. For travel journalism, real-estate or a news context, say that the image is edited. For your own album, nothing more is needed. See mask checklist and limits for the file rules.