How to keep only a few objects and delete the rest of the background
Name the objects you want kept and ask Vizard Agent for a transparent version. Automatic background removal is trained to find people, so on a shot with no person in it the result is unpredictable — the reliable route is a mask built around the objects you listed, checked against the alpha channel.
What is the short version?
You have a clip of a rotating cake with a candle on a desk, and you want those three things on transparency so you can put them over something else. The automatic route is built for a different job. Vizard Agent builds the mask deliberately instead.
- Tell Vizard Agent exactly which objects stay.
- Expect a mask rather than a one-pass cutout.
- Ask to see the alpha channel, not just a frame.
What do you need before you start?
The clip, and a clear list. "Remove the background" is ambiguous on a shot full of things; "keep the cake, the candle and the desk, remove everything else" is a specification Vizard Agent can build against and check afterwards. The list is also what makes the review meaningful, because each named object either survived or did not.
It helps to know where the objects move. A cake that rotates changes its silhouette every frame, so the mask has to follow it rather than sit still. Vizard Agent samples the clip at several points to see the range of motion before it fixes the mask boundaries, and telling it what moves saves a pass.
What do you type into Vizard Agent?
List the keepers. The instinct is to describe what should go, and that reads naturally but gives the tool nothing to anchor on — "remove the background" only means something once you have said what the foreground is, and on a busy shot that is not obvious.
Keep the cake, the candle and the desk. Delete everything else and give me a transparent video.
That was the brief behind this article, and it worked because it was a list rather than a description. Vizard Agent could check each item against the result and report on all three separately rather than making an overall judgement about whether it looked right.
What does Vizard Agent actually do?
It tries the automatic route first, looks at what came back, and switches to a built mask when the automatic pass keeps the wrong things. That sounds inefficient and is not — the automatic pass is fast, and on the occasions it works it saves the rest of the job.
In that session the sequence ran:
- Pulled frames across the whole clip to see what the foreground actually consists of.
- Found the video's display rotation lived in metadata rather than in the pixels, so the cutout was being fed a sideways picture.
- Baked the rotation into the frames and re-ran the removal on a correctly oriented file.
- Built a custom alpha mask around the three named objects when the automatic result still kept the wrong foreground.
- Tightened it twice more, first for general background residue and then specifically around the cake.
- Reviewed the alpha channel itself, and a preview of the foreground over grey, rather than trusting a frame that looks fine against black.
That rotation step is worth noting. A clip that plays upright in every player can still be stored sideways with a rotation flag, and a tool that reads the pixels sees the sideways version.
What does the result look like?
A video with a real alpha channel, containing the objects you named and nothing else, that you can composite over anything. Vizard Agent checks it over a grey background, because transparency against black looks identical to black pixels and most bad cutouts pass that test.
You also get to see the mask on its own. A mask that looks reasonable and a result that looks reasonable are different checks, and residue around a moving edge is much easier to spot in the mask than in the finished frame.
When does this not work well?
When the objects you want touch the edge of the frame. A desk that runs off both sides has no outline there, so the mask has to guess where it ends, and the guess is usually visible once you composite it onto something with a different colour.
It also degrades when the objects are the same tone as what is behind them. A dark desk against a dark wall gives the mask very little to find, and each tightening pass that removes wall also removes desk. Vizard Agent will say when it is trading one for the other rather than quietly picking a side.
And it is harder on fast motion than on rotation. A steadily turning cake presents a predictable range of silhouettes; something whipping through the frame with motion blur has edges that are genuinely ambiguous, and no mask resolves them cleanly.
How do you fix a result that came back wrong?
Say which object is wrong and how. "The candle is gone" and "there is still wall around the cake" lead to different corrections, and Vizard Agent can act on either immediately. In that session the tightening happened in exactly that order — general residue first, then a pass aimed only at the cake.
If everything came out at the wrong angle, that is the rotation problem rather than the mask. Ask Vizard Agent to check the rotation metadata against the actual frame orientation, which is a different question from whether the cutout worked.
If the file looks solid rather than transparent when you open it, the alpha may not have survived the encode. Ask Vizard Agent to verify the channel in the delivered file, because a preview showing a checkerboard is not proof that the file carries one.
How does Vizard Agent compare to doing it yourself?
Rotoscoping three objects across a clip by hand is genuinely slow work, and it is the reason automatic cutout tools exist. But those tools all share the same bias: they are trained on people, and on a shot with no person the result is not merely worse, it is aimed at the wrong thing entirely.
What Vizard Agent adds is the check rather than the mask. You would build a mask too; you would probably also look at it over black, decide it was fine, and find the residue when you composited it. Reviewing the alpha over grey and inspecting the mask separately is the part that gets skipped by hand and the part that catches the problem.
Common questions
Why did the automatic tool keep the wrong things? Because it looks for a person. Vizard Agent switches to a built mask when there is not one.
Can I keep a person and an object? Yes. Name both and Vizard Agent masks them together.
What format is the transparent file? A codec that carries an alpha channel. Vizard Agent picks one that survives the export rather than the smallest file.
Will it work on a still image? Yes, and more easily. A single frame gives Vizard Agent one silhouette rather than a moving one.
How many objects can I name? As many as are genuinely distinguishable. Vizard Agent will tell you when two of them are effectively one region.
Does the mask follow movement? Yes. Vizard Agent samples across the clip so the mask covers the object's full range of motion.
Can I see the mask itself? Ask for it. Vizard Agent can output the alpha channel as its own video.
Why check it over grey? Because transparency over black is invisible. Grey shows both the holes and the residue.
What if the edges look hard? Ask for feathering. Vizard Agent can soften the boundary so it composites without a cut-out look.
Does this change the timing? No. The clip keeps its original length and frame rate.
Can I composite it in the same job? Yes. Tell Vizard Agent what goes behind and it delivers the finished composite as well.
What about shadows the objects cast? They are background unless you say otherwise. Tell Vizard Agent to keep the shadow if you want it.
Does the audio survive? Yes, unless you ask for it to be dropped.
Will a second pass lose quality? No. Vizard Agent rebuilds the mask from the source rather than re-processing the previous result.