A person from one photo, standing in another
The friend who missed the group photo, the family member in another city, the portrait that deserves a better background: the most common reason to merge two photos is to put a person from one into the other. That needs the person’s outline, and drawing it by hand around hair and fingers takes an afternoon. A small model finds it in about a second — and on this site the model runs on your own phone or computer, so neither photo has to be handed to a service to do it.
The “AI” here is that outline and nothing more. It does not paint, invent or change anyone. The rest is ordinary photo arithmetic you can see and undo: the cut-out is scaled, placed, nudged toward the scene’s colour by an amount you choose, softened at the edge and given a shadow.
What happens, step by step
- Finding the person. PP-HumanSeg (Apache-2.0) looks at a 192 × 192 copy of the first photo and scores every point for how likely it is to be a person. That map is stretched back to the photo’s shape with a narrow soft edge.
- The first download. Before anything downloads you see the sizes — a 5.9 MB model and a 13.4 MB runtime — and choose to go ahead. The model is checked against its recorded fingerprint before it runs, then kept in your browser so the next time is instant.
- Placing. The person’s height is a share of the new photo’s height, their feet stand at a point you choose, and they can be mirrored to face the other way. Drag them on the preview to move them.
- Blending in. Match the light moves the person’s average colour toward the colour of the place they now stand; the list under the preview prints the change in stops for red, green and blue.
A worked example
A 3000 × 4000 portrait of a cousin against a white wall, and a 4000 × 3000 beach photo. The model marks about 30 % of the portrait as person. At the default 70 % height the cousin is placed 2100 pixels tall, feet 98 % of the way down the beach photo. Match the light at 50 % warms them toward the sand by a fraction of a stop in each channel; Edge in at −2 removes the last trace of white wall around the hair. The result is saved at 4000 × 3000.
Privacy, specifically
Your photos stay in this tab from start to finish. Turning on the person finder downloads the model from our model file server once; that request carries no picture and no information about it. Nothing is generated by a remote service, and nothing about your photos is kept after you close the tab.
The model finds people only, and works best with one person clearly in the frame; groups come out as one shape. It cannot relight, so photos with very different light still look different. The outline cannot be painted by hand here — only moved in or out and softened. On older phones the first run can take several seconds.
Questions about putting a person into another photo
Is my photo uploaded to an AI service?
No. The model — a 5.9 MB person-outline finder — is downloaded to your browser once, from our model file server, and runs in this tab. Your photos are never sent there or anywhere else; the download goes one way, from us to you. After the first time it is kept in your browser, and the privacy page has a button that deletes it.
Does this generate a new picture?
No, and that is deliberate. Nothing is invented: the person is cut out of the first photo along the outline the model finds, and placed into the second. Every pixel of the person comes from your first photo and every pixel around them from the second. It will not change a face, a pose or clothing, and it will not make a person who was not photographed.
The cut-out has a halo of the old background. How do I fix it?
Move Edge in or out a few pixels to the left; the outline pulls in and the halo goes. If hair or fingers get clipped instead, move it the other way. Soft edge blends the outline into the new scene; 2 to 4 pixels usually looks natural, more looks dreamy.
Why does the person look pasted on?
Usually light and size. Raise Match the light so the person's colour moves toward the scene's; set the height so they are the size a person would be at that spot; stand their feet on the ground rather than in the air; and keep the soft shadow on. The model cannot relight a face lit from the left for a scene lit from the right — choose photos with similar light.
It says no person was found.
The model looks for people, not pets, products or statues, and works best when the person is reasonably large in the frame. Crop the photo closer, or use one where the person stands out from the background. For a logo or product on a plain background, the photo-on-photo tool's flat-background removal works without any model.
