Remove a person from a photograph
A wider brush is ready and the structure-aware option is selected, but its model is not fetched until you accept the disclosed download. Patch search remains one click away.
Drop in a photograph and paint over whatever shouldn't be in it.Open the highest-quality authorised source you have. The mask and fill stay on this device.
Nothing to hand?
The opening state is a starting judgement
A person is usually a large, irregular hole rather than a narrow defect. This route therefore begins with a larger brush and points at the structure-aware fill, while still requiring explicit consent before any model weights arrive.
The selected mask instrument and fill method are visible and can be changed. Draw slightly beyond the visible object so shadow, antialiasing and colour spill are included, then remove any legitimate edge the mask crossed before running the fill.
A disciplined removal pass
- Mark the complete footprint. Include the object, its soft shadow and any reflected colour that belongs to it.
- Choose from the surrounding evidence. Use a narrow march for a stroke, patch search for texture and the disclosed model only when the hole needs structure.
- Read the divider at one-to-one. Look for repeated texture, bent straight lines and a smooth island where natural grain should continue.
- No fill reveals a hidden original; every new pixel is inferred from visible context.
- A unique face, sign or architectural edge behind the mask cannot be recovered exactly.
- Several small masks are usually safer than one huge region crossing unrelated surfaces.
Questions about the repair
- What pixels are changed?
- Only the mask and its feathered edge. The rest of the working frame remains untouched; a full-size export reruns the same operation against the source dimensions.
- Which fill should I trust?
- Edge march suits very narrow marks, patch search suits repeated natural texture, and the optional structure-aware model suits a larger hole. Each fails differently, so the before-and-after divider matters more than the label.
- Does the model receive my photograph?
- No. If selected, its disclosed weights are downloaded to the browser after consent. Inference then runs on this device and no image request is made.