How to catch b-roll that shows something similar but not the thing you named

Last updated 2026-09-10 · 8 min read

What is the short version?

Ask for a shot of one thing and you will often get a shot of something that resembles it. It passes every check that asks "is this food being sliced" and fails the only check that matters, which is whether it is the food you named.

1. Go to Vizard Agent with the subject named precisely.
2. Ask it to confirm what each clip actually shows.
3. Ask it to generate the shot if searching keeps missing.

What do you need before you start?

The subject, stated as specifically as you would to somebody who knows the field. Vizard Agent searches on what you give it, and a category — "food", "machinery", "a bird" — returns whatever the library thinks fits that category, which is where the near misses come from.

• The exact subject. The dish, the species, the model.
• What makes it identifiable. Colour, shape, context.
• Whether a near match is acceptable. Sometimes it is.
• Who will notice. An expert audience notices everything.
• Whether you can generate. It is the fallback.

What do you type into Vizard Agent?

Name what would make a clip wrong, not just what would make it right. Saying which near neighbour you do not want is more useful than another adjective, because that is exactly what the library is going to offer.

Prompt

Variants worth knowing:

• "Not pork, not generic roast meat." Rules out the near miss.
• "Check what each clip actually shows." Verification, not search.
• "Generate it instead." Authorises the escalation.

What does Vizard Agent actually do?

Here is how it went on a real ten-second opener that needed one very specific dish. The first search returned a perfectly plausible clip, the user spotted that it was the wrong thing, and the useful part of the story is what Vizard Agent did next.

1. Researches the subject so it knows what it is looking for.
2. Searches stock for the specific shot.
3. Downloads the candidates.
4. Looks at frames from each rather than reading the titles.
5. Compares what is shown against what was asked for.
6. Searches again with a narrower term when it does not match.
7. Generates the shot when the library keeps missing.
8. Re-assembles with the correct footage.
9. Reviews the finished frames to confirm the fix landed.

Step four is where near misses get caught. Stock titles and tags describe the clip generously, and a clip called "chef slicing meat" is exactly as true of the wrong meat as the right one — only looking at the frames settles it.

Step seven is the escalation people forget they have. Two failed searches are usually a sign the library does not hold the shot, and continuing to search is cheaper per attempt and more expensive overall than generating the specific thing once.

Step one earns its place on unfamiliar subjects. Knowing what the thing actually looks like is what makes step five possible — without it, a check can only confirm that the clip contains something in the right category.

What does the result look like?

Footage that shows the thing you were actually talking about, along with a note from Vizard Agent of which clips were replaced and why each one failed. On any subject your audience knows well, that is the difference between a video they trust and one they correct in the comments.

When does this not work well?

Some shots do not exist and cannot be convincingly generated. A rare object, a specific real place, or a real person doing a particular thing are all cases where the honest answer is to change the line rather than to keep hunting for the picture.

• Rare or specific real subjects. They may not exist as footage.
• Real people or places. Generation is not the answer.
• Fast cuts. A near miss may genuinely not matter.
• Abstract lines. There is no literal subject to match.
• Very specialist fields. Ask someone who knows to check.

How do you fix a result that came back wrong?

Say what the clip actually shows rather than that it is wrong. "That is pork, not duck" tells Vizard Agent both what the mistake was and what to exclude on the next attempt, whereas "that clip is wrong" simply starts the same search over again.

• "That is pork, not duck." Excluded and searched again.
• "Still not right — generate it." Straight to the fallback.
• "That one is fine, keep it." A near match you accept.
• "Wrong part of the process." Carving, not roasting.

How does Vizard Agent compare to doing it yourself?

By hand the search results get skimmed as thumbnails and the closest-looking one goes in. Nobody looks at it again at full size, and the mistake is found by a viewer who knows the subject better than anybody who made the video.

Common questions

Why do libraries return the wrong thing? Because they match the shape of the request rather than its subject.

How many searches before generating? Usually two. Vizard Agent will suggest the switch rather than looping.

Will a generated shot look right? Often, for food and objects. Vizard Agent will say when it is risky.

Can it check clips I supplied? Yes, the same way — by looking at what they show.

What if I do not know the subject well? Vizard Agent researches it first, which is what makes the check possible.

Does the clip title help? Barely. Vizard Agent treats titles as a hint, not evidence.

Can it list what it rejected? Yes. Vizard Agent reports each rejected clip with the reason.

What about people? Generation is the wrong tool there; Vizard Agent will say so.

Is a near miss ever fine? Yes, in a fast cut. Tell Vizard Agent when you are happy to accept one.

Will it slow the job down? A few minutes of Vizard Agent checking, against a correction after publishing.

Can it check a whole edit at once? Yes. Ask Vizard Agent to sweep every b-roll shot against its line.

Does this apply to generated b-roll too? Yes. Vizard Agent checks generated footage the same way; it can also be wrong.

What if the shot does not exist? Vizard Agent will tell you, so the line can change instead.

Why did nobody notice? Because everyone who made the video was checking the category rather than the subject itself.