How to remove a person from a video and keep what they were holding

Last updated 2026-08-16 · 4 min read

What is the short version?

The instruction that works is not "remove the person". It is "remove the person, keep the remote control in their hand and the toy on the floor". Removal is easy to describe and easy to over-apply, and everything you fail to mention is at risk.

1. Go to Vizard Agent and upload the clip.
2. Say what to remove, in one specific phrase.
3. List what must stay untouched, especially anything being held.

What do you need before you start?

A short clip, and two lists: what goes, and what stays. The second list is the one people skip, and it is the one that decides whether the result is usable — Vizard Agent removes what you name, and anything you did not think to protect is inside the region it is working on.

• A locked-off camera, if you can. A tripod shot is far easier than a handheld one, because the background behind the person is the same in every frame.
• A few seconds, not a few minutes. Removal is per-frame work. Start with the shot that matters.
• A frame where the thing you want to keep is fully visible. If the object is hidden behind the person for the whole clip, there is nothing to keep.
• The list of exceptions. Write it as a list. "Remove the person. Do not remove the remote control. Do not remove the vehicle."

What do you type into Vizard Agent?

Write it as an instruction plus its exceptions, in that order. Vizard Agent honours negative clauses well, so "do NOT remove the remote control" is a real constraint rather than a hint — and stating it costs one sentence against a second pass that costs a whole run.

Prompt

Variants:

• Remove a fixed object. A sign, a bin, a logo on a wall. Easier than a person, because it does not move.
• Remove someone in the background. Say where they are and when they enter, which saves Vizard Agent finding them.

What does Vizard Agent actually do?

Here is the order it worked in on a real clip. The method is worth reading, because it explains the limits — every one of them follows from Vizard Agent having to rebuild what it cannot see. This is one run, not a typical one; the shape holds, the effort varies:

1. Extracts frames across the whole clip and looks at them to see what is where.
2. Tries the video editor directly on the removal.
3. Measures camera stability across the video with OpenCV. This decides the whole approach: a still camera means one clean background can serve every frame.
4. Takes one full-resolution frame and inpaints it with the image editor, rebuilding what is behind the person.
5. Aligns the rebuilt background back to the original geometry and makes a side-by-side comparison to check the alignment.
6. Draws a coordinate grid over the original frame and crops to the region around the person and the objects, so the boundary between "goes" and "stays" is worked out on real pixel coordinates rather than by description.

Step 3 is the hinge. Vizard Agent measures the camera before choosing a method rather than assuming one.

What does the result look like?

The same shot with one thing gone and everything you protected untouched. Vizard Agent keeps the clip, the extracted frames and the rebuilt background plate in the project, so a second pass at a stubborn region reuses all of that rather than starting the reconstruction again.

From this run, probed on the delivered file: 720x1280, H.264, 30fps, 5.23 seconds, audio preserved (AAC). The output followed the source framing here; where a destination needs something different, say so and it will be delivered that way.

This category was not separately measured. Comparable work runs a median of 28 to 38 minutes end to end, and across all projects the median cost by tier is Flash 47, Pro 55, Max 242, Ultra 263 credits.

When does this not work well?

Removal is reconstruction, not erasure — whatever was behind the person has to be rebuilt from somewhere. That single fact explains every limit below, and it is why Vizard Agent measures the camera first: a still camera means one honest plate, and a moving one means guessing again every frame.

• A moving camera makes this much harder. With a handheld shot the background shifts every frame, so a single rebuilt plate no longer fits. It can still be done, and it will take longer and hold up less well.
• Whatever was fully hidden cannot be recovered. The rebuilt background is generated, not uncovered. If the person stood in front of a face, a sign, or a detailed pattern for the entire clip, what appears there is invented, plausibly.
• Shadows and reflections are part of the person. A person removed from a sunlit floor can leave their shadow behind. Say "and their shadow" or expect to ask twice.
• Contact points are the hardest pixels in the shot. Fingers around a remote, a hand on a handle. That boundary is exactly where the removal has to be surgical, and it is where a second pass usually goes.
• This is not a legal opinion. Taking a recognisable person out of footage does not settle whether you had the right to film or publish them.

How do you fix a result that came back wrong?

Reply in the same Vizard Agent conversation instead of starting over. The clip, the extracted frames and the rebuilt plate all stay in the project, so a correction is a second pass over a specific region rather than a fresh reconstruction of the whole shot.

How does Vizard Agent compare to doing it yourself?

By hand this is a VFX task: mask the subject on every frame, clean-plate the background, track the plate to the camera, then composite. Rotoscoping alone is measured in hours per second of footage, which is why routine removals simply do not get done.

Neither replaces the other. Vizard Agent wins on shots where the removal is a chore rather than a craft; a VFX artist still wins on a moving camera over a detailed background, which is exactly where the reconstruction gets hardest.

Common questions

Can it remove several things at once?

Yes. List them, and list what stays. The list matters more as the number grows, because everything unnamed sits inside the region Vizard Agent is reconstructing and is therefore at risk.

Will it damage the rest of the picture?

Only the region it works on changes. That is why naming what stays is worth the extra sentence — it tells Vizard Agent where the boundary is instead of leaving it to be inferred from a description.

How long a clip can it handle?

Give Vizard Agent the shot, not the whole take. Removal cost scales with frames, so a five-second clip is a very different job from five minutes, and a first pass on the shot that matters tells you whether the approach holds.

What if the result looks smeared where the person was?

Say where. Vizard Agent still has the clip, the frames and the rebuilt plate, so the next pass corrects that region instead of starting over — and being specific about which part of the frame is what makes that cheap.

Does it work on a handheld shot?

It can, and it holds up less well. Vizard Agent measures camera stability before choosing a method, so tell it the shot moves and expect the reconstruction to be softer than a tripod shot would give you.