Remove a Photobomber from a Photo — Pro, on our server
Occlusion runs the right way round, and that is the whole reason this works. Every removal is a question about missing information. When somebody stands beside your subject with an arm across them, part of your subject was never recorded, and removing the intruder forces the model to author a person — the difficult territory described on removing someone who is touching you. A photobomber stands behind, so the occlusion is the other way about: your subjects hide parts of the stranger, and nothing belonging to the people you are keeping is absent. Mask the visible portion and the model has one job, background reconstruction, at which it is genuinely good.
Depth of field is quietly doing you a favour. Anything focused on a group a few metres away throws the space behind them out of focus, and a defocused background is a soft, low-detail wash with no sharp edges to align and no fine texture to match. That is precisely the material inpainting reproduces most convincingly, because the fill inherits the character of what surrounds it — a blurred mask region is filled with blur. A stranger twenty metres back in front of a defocused hedge or a soft wall tends to disappear so completely that there is nothing left to inspect. The corollary is that a photobomber in the same focal plane as your group, sharp and detailed, is a different and harder job.
Mask to the edge of the people in front, and no further. The instinct is to paint over your friend’s head to “get all of” the stranger behind it. Do not — everything under the brush is replaced by generated pixels, so painting over your friend replaces part of your friend. Follow her hairline with a small brush, stop there, and let the fill handle only what the camera actually saw of the intruder. Do include the extras people forget: a hand raised above a shoulder, a hat, a bright bag strap, the slice of a leg visible between two subjects. Then run once and check along that boundary at full size, because a wobble in the mask along a hairline is the one artefact this otherwise easy job produces.
Cheaper answers first, and the terms of the run. Look through the burst for a frame without them; look at whether a crop takes them out along with a corner you never liked. Both are free, on-device, and reconstruct nothing. If neither works, the fill is Cloud HD — Pro-only, opt-in, run on our GPU in the United States, with the photo and mask encrypted in transit, deleted when the run finishes and in all cases within 24 hours, and never used to train anything. Pro includes 150 runs a month on Monthly and 200 on Annual, of which 10 may be HD. Related cases live on person removal, crowds at a landmark and unwanted objects generally, all of them the same object remover with a differently shaped mask. If the group photo also needs sharpening or noise cleaned up, that is a whole-frame job for the photo enhancer, not a mask.