Vizard Agent

How to clean marks off a surface in a video

Last updated 2026-08-30 · 7 min read

Upload the clip and tell Vizard Agent which surface needs cleaning. Vizard Agent crops in to see the marks properly, computes the motion across the whole clip to find where the surface stays still, confirms the camera has not drifted, and removes the marks without touching what belongs there.

What is the short version?

Cleaning marks off a mirror, a wall or a bonnet in footage is not a paint job, it is a tracking problem. If the surface holds still relative to the camera it is straightforward; if it does not, every frame is a new problem.

  1. Go to Vizard Agent and upload the clip.
  2. Say which surface and what should come off it.
  3. Say what must stay — reflections, fittings, genuine detail.

What do you need before you start?

The clip and a clear boundary. Vizard Agent measures the camera movement itself, so you do not need to know whether the shot is locked off, but you do need to say what counts as dirt and what counts as the surface.

What do you type into Vizard Agent?

Say it plainly. "Make the mirror clean" is a complete instruction, and adding what must not change — the reflection, a maker's mark, a scratch that is part of the object — is what stops the cleaning going further than you meant.

Prompt

Make the [mirror] clean — remove the smudges and spots. Keep the reflection and the frame exactly as they are.

Variants worth knowing:

What does Vizard Agent actually do?

Here is the order Vizard Agent worked in on a real surface-cleaning job. The stretch from the fifteenth step is the real work: measuring how much of the frame actually holds still, because that determines whether one clean patch can serve the whole clip.

  1. Examines the footage and extracts sample frames, then builds a preview grid.
  2. Inspects the frames to understand what needs cleaning.
  3. Checks the audio stream and whether there is a speech transcript.
  4. Crops close-ups of the surface to inspect the dirt and the stains properly.
  5. Views those close-ups, then the surface across the full frame.
  6. Views the region containing the surface at several points in the clip.
  7. Checks the video's dimensions and analyses the camera's stability.
  8. Surveys the shot structure and checks the available capabilities.
  9. Computes the motion variance across the whole clip to locate the static regions, inspects the motion map, tracks the surface's movement across every frame, estimates the camera motion, and compares frame zero against frame three hundred to confirm.

Step 9's last check is the one that decides the method. If the frame at ten seconds matches the frame at the start, a single cleaned patch can be held across the clip; if it does not, the patch has to move with the surface and the job is several times larger.

What does the result look like?

From the run this page is written from, probed on the delivered file: 1920x1440, H.264, 30fps, 20 seconds, AAC audio. The same twenty-second clip with the surface clean — the smudges and spots gone, the reflection intact, the frame and the fittings untouched, and the audio unchanged.

Nothing else in the shot moved. That is the requirement for this kind of job: a viewer should not be able to tell that anything was done, only that the surface looks as it should have on the day.

When does this not work well?

Removing marks means replacing them with whatever should have been there instead, and that gets steadily harder as the surface or the camera starts moving. These are the cases where Vizard Agent runs into the footage rather than the instruction.

How do you fix a result that came back wrong?

Point at the frame where it fails and say what you can see. Vizard Agent keeps the close-up crops, the motion map, all the tracking data and the stability measurements, so the patch can be re-fitted without re-analysing the clip.

How does Vizard Agent compare to doing it yourself?

By hand this means masking the surface, cloning clean pixels over the marks, and then discovering the patch slides as soon as the camera breathes. On a handheld shot most people give up and reshoot with a clean mirror.

By hand Vizard Agent
Camera movement Discover it during the fix Measured across the whole clip first
The marks Eyeball them at normal size Cropped and inspected close up
Method One approach, tried Chosen from the measured motion
Verification Watch it back Frames hundreds apart compared

Common questions

Does the camera need to be locked off? No, but it is easier. Vizard Agent measures the motion and picks its method accordingly.

Can it clean a mirror without losing the reflection? Yes. Vizard Agent removes the marks and leaves whatever the mirror is showing intact.

What surfaces does this work on? Glass, paintwork, walls, screens. Vizard Agent handles anything where marks sit on something uniform.

Will it remove marks I wanted to keep? Not if you name them. Otherwise Vizard Agent treats every mark as a candidate.

What if the whole surface is dirty? Then replacing it may be cleaner than patching. Vizard Agent will say which is better.

Why compare frames hundreds apart? Because a shot that looks locked off can drift slowly enough that nothing is obvious in playback, and that drift is exactly what makes a held patch slide off the surface halfway through.

Does the audio change? No. Vizard Agent leaves the sound exactly as it is.

Can it clean several surfaces at once? Yes, though Vizard Agent tracks each of them separately.

How long can the clip be? Any length. Longer clips mean more frames to track.

Is it better to clean the surface before filming? Always, where you can. Vizard Agent does this well on a locked-off shot, but a wiped mirror and a reshoot costs nothing and produces a result nobody has to inspect afterwards.

Can it clean a surface that only appears sometimes? Yes. Vizard Agent surveys the shot structure first, so it knows which stretches contain the surface at all and only works on those rather than processing the whole clip blindly.

Does Vizard Agent check the result? Yes. It reviews the cleaned surface across the clip before delivery.