B-roll

AI B-roll generator, with the result shown

B-roll is the footage you cut away to while the main audio keeps running. In a talking head video it does three jobs: it marks a change of subject, it hides a bad cut, and it gives the viewer somewhere else to look before they get bored of your face. An AI B-roll generator automates finding or making those cutaways. The rest of this page is about what that actually produces.

The same cut, twice

Same audio, same take, same length. The only difference is three generated inserts placed at the topic changes. Watch them one after the other rather than side by side, because the effect is about pacing and pacing does not survive a split screen.

Two products, one name

Before you pick a tool, work out which of these it is. The marketing copy is nearly identical and the workflow is not.

What generated B-roll is good at

Generation wins where a library is thin and where the shot only has to feel right rather than be a real place.

• Abstract and macro. Ink, circuitry, fabric, light through glass. High detail, no identity to preserve, no failure modes.
• Faceless human activity. Hands typing, hands writing, someone walking away from camera. All the usefulness of a person with none of the face problems.
• Things that do not exist. Your specific product, a diagram in motion, a scene that no crew ever shot.
• Matching a look. If your main footage is dark and cool, you can prompt for dark and cool. A library gives you what it has.

Where to place inserts

1. At every topic change

   The cut tells the viewer something new has started. This is the one that helps the viewer rather than the editor.

2. Over the joins

   If you cut two takes together and the jump is visible, an insert across the join hides it and the audio carries straight through.

3. Over your weakest thirty seconds

   Every talking head has a stretch where the energy dips. That is where a cutaway earns its generation cost.

4. Not at the open

   Start on you. Cutting away in the first five seconds tells the viewer they are watching a slideshow, and they leave.

Prompting for B-roll specifically

B-roll prompts are different from hero shot prompts. You want them boring, loopable and short, because they run under speech and nothing in them should pull attention off the voice.

• Ask for one slow motion, not an action. Drift, not event.
• Keep faces out. Hands, backs of heads, and out of focus people all work.
• Say no text explicitly. Signs and screens are the most common ruin of a usable insert.
• Match your main footage in three words: the light direction, the colour temperature, and the lens. Repeat those three across every insert.
• Generate at four to five seconds and cut to two or three. The end of a clip is the worst part.

The workflow this belongs to

B-roll is the second half of a script driven edit. The first half, turning writing into scenes and prompts, is on the script to video page. If you are making longer content where cutaways matter most, the tool comparison for that is the YouTube page. For the wider picture of what these models do, start at the video AI generator overview.

Twelve B-roll prompts that keep working

These are deliberately generic. B-roll should not be interesting on its own, it should be unremarkable enough that the viewer keeps listening. Swap the nouns and keep the structure.

1. Slow macro drift across a textured surface, shallow depth of field, no text, four seconds.
2. Hands typing on a laptop, over the shoulder, static camera, natural window light, face out of frame.
3. Steam rising from a cup on a desk, static camera, warm side light, nothing else moves.
4. A pen moving across a notebook page, top down, hands only, soft light, three seconds.
5. Light moving across a wall as a cloud passes, static camera, no people, five seconds.
6. An empty office corridor, slow dolly forward, cool light, no signage, four seconds.
7. Rain on a window at night, static camera, blurred lights behind, no reflections of people.
8. A city street from above at dusk, slow pan right, traffic moving, no readable signs.
9. Dust in a shaft of light, macro, static camera, dark background, four seconds.
10. A door closing slowly in an empty room, static camera, side light, no hands.
11. Ink spreading in water, macro, black background, no camera movement, four seconds.
12. Someone walking away from camera down a path, back view only, overcast light, five seconds.

Every one of those avoids faces and text, asks for one motion, and names the camera. Those three habits are what makes a prompt usable rather than clever.

Making generated inserts sit in a real edit

1. Match the frame rate at import

   Generated clips arrive at whatever the model produces. Convert them to your timeline rate before you place them, not after, or every insert stutters at its own cut.

2. Match the colour

   Generated footage has its own contrast and cast. A quick match to your main footage does more for the cut than a better prompt would have. This is the step people skip and it is the one that gives the insert away.

3. Trim to the middle

   Generate four to five seconds and use the middle two or three. The first frames often have a small settle and the last ones have the most drift.

4. Keep the audio underneath

   The point of B-roll is that the voice continues. Never cut the audio at the same moment as the picture unless you want the viewer to feel the join.

5. Add a little motion in the edit

   A one or two percent scale ramp across the insert helps it feel like part of the same edit as your handheld footage. More than that reads as an effect.

Cost, in the unit that matters

B-roll is the cheapest kind of generation to plan for, because the subjects are easy and the failure rate is low. On texture, abstract and faceless human activity, most takes are usable, which is not true anywhere else in video generation.

When to use stock instead

Generation is not always the right call and pretending otherwise would make this page as useless as the ones it is competing with. Use a library when the shot is a real recognisable place, when you need a long continuous take, when the shot contains readable text, or when the subject is people doing something specific with their hands and faces. All four are things generation currently does badly and a camera crew already did well.