How to use AI B-roll in a talking-head video without it looking like stock

A talking-head video is one shot held for ten minutes. However good the speaker, the eye wants somewhere else to go, and the edit's job is to give it one without interrupting the voice. That is B-roll: everything you cut to while the person keeps talking. In editors' shorthand, the person talking is the A-roll.

B-roll used to mean filming it yourself or paying for stock. Now you can generate exactly the shot a sentence describes, in a minute or two, for less than most stock licences. That is the opportunity. The trap is that generated B-roll which illustrates every sentence literally turns a video into a stock montage assembled by someone who never met the speaker.

Cut the talk first

B-roll goes on top of a finished spine, never before it. Tighten the recording first: cut the silences and the filler words (both scans are in the clip's Audio tab, and the long version is here), take out the false starts, and settle the order of the points. Every cut moves everything after it, so B-roll placed too early ends up over the wrong sentence.

Tightening has a second use. Every cut you just made is now a visible jump in the speaker's face, and each jump needs covering one way or another: with B-roll, or with a punch-in, where the clip after the cut is framed about 15% tighter (Transform → Crop) so the jump reads as a second camera. The punch-in is the right fix wherever the face has to stay on screen.

Put it only where it has a job

B-roll has three jobs. An insert that is not doing one of them is decoration, and decoration is what makes a video feel padded.

1. Hide a cut. A two-second insert centred on the join removes the jolt, and the voice runs straight through.
2. Show the thing. When the speaker describes something the viewer cannot see (a place, an object, a process, a before and after), show it, as close as possible to the moment it is named.
3. Rest the eye. After a long unbroken stretch of face, a cutaway resets attention. How long is too long depends on the format: well under twenty seconds in a short, a minute or more in a calm long-form video.

Just as important is where it does not go: on the moment of feeling, on the punchline, and on the line said straight to camera. There, the face is the point. Covering it with a shot of a laptop throws away the one thing the speaker brought.

Show the thing, not the idea

This is the difference between B-roll that helps and B-roll that looks like a screensaver.

When the speaker says "our bookings tripled", the stock instinct is a glowing graph with an arrow, or two people shaking hands. That illustrates an idea, success, and the viewer has seen it a thousand times. Show the thing instead: the actual café, the actual product on a shelf, the diary that is now full. Or put the number on screen as a title and skip the picture. Concrete beats symbolic, every time.

A few rules keep generated inserts honest:

• Match the register of the A-roll. If the speaker was filmed on a phone at a kitchen table, the B-roll should look captured too: handheld, natural window light, ordinary places. Golden-hour cinematic inserts under a webcam talking head are the mismatch viewers notice first. Put it in the prompt: "handheld phone footage, overcast daylight, an ordinary office kitchen".
• Leave faces and hands out where you can. They are where generation goes wrong and where viewers look hardest. Places, objects, details and wide shots are far safer.
• Keep it short. Two to four seconds an insert. Generated shots are at their cleanest early on, and B-roll rarely needs to be longer anyway.
• Keep the camera calm. A locked-off frame or a slow push. A busy move competes with the voice. What each camera move says covers the choice.
• Film the real things. The speaker's actual shop, product or street cannot be prompted into existence, and a phone clip of it beats any generation. If all you have is a photo, animate it (Generate → Video → Animate), and treat the result as AI footage when you publish.

Put it on a lane above

The mechanics take a minute. Add a video lane above the talking head with the + beside the lane headers. A higher lane draws over the lanes beneath, so drop each insert where it belongs and the speaker keeps playing underneath it, voice and all. Nothing is cut out of the A-roll, which means you can slide an insert a few frames either way, or delete it, without touching the talk. Layers and picture-in-picture has the detail.

Then three things:

• Silence the insert. Some video models generate their own sound. Select the insert, open its Audio tab and press the speaker icon beside Clip volume, or turn the model's sound off before generating.
• Land it on the word. Start the insert on the noun it shows, or a few frames before, so the picture arrives as the word does. End it at the end of a sentence, never halfway through a word.
• Keep the captions on top. Text and captions live on the overlay track above every video lane, so they read straight across the inserts without moving.

Make it match

The giveaway with generated B-roll is rarely what is in it. It is the finish: inserts more saturated and more contrasty than the face they cut away from, and often softer, because they were generated at a lower resolution. Three passes fix most of it.

Colour. Grade the inserts toward the A-roll, not the other way round. Effects → Color has temperature, saturation and contrast per clip, and generated footage usually wants a little less of the last two. Then put one LUT across everything, so the whole video looks like it came from one camera. LUTs explained has the order.

Resolution. Generated clips often come out at 720p. Upscale the inserts you keep, 2×, so they hold up at the export size next to footage from a real camera.

Shape. Generate inserts in the project's own aspect, 9:16 for a short and 16:9 for YouTube, rather than cropping a wide shot into a tall one and losing most of it.

How much is enough

For a typical explainer, B-roll covering somewhere between a fifth and two fifths of the running time is a good range. For an opinion piece or a personal story, where the face is the product, less. Past half, it has stopped being a talking-head video and become a narrated montage, which is a different thing and wants a different script.

One rule for the upload

Realistic generated footage counts as AI content on YouTube, TikTok and Instagram, B-roll included, so tick each platform's AI disclosure when you upload. YouTube says the label does not, on its own, limit a video's reach.

The test

Watch the finished cut with the sound off. If the inserts alone tell you roughly what the video is about, the B-roll is doing its job. If they could belong to any video on any subject, they are illustrating ideas, and it is worth one more pass with a list of the actual things the speaker mentions.