How AI matches b-roll to your script
By Neil, founder of Lantaw. Last updated: 2026-11-11
Lantaw matches b-roll to your narration by cutting the script into short scenes, having an AI scene director write one shot for each scene, and generating a new clip from every shot. It does not search a stock library for keywords. Each clip is cut to the length of the narration it covers, so the picture follows the voice.
This page explains each step, the rules the director follows, how to write lines that are easy to picture, and where the footage can still go wrong. It describes how Lantaw works, not how any other tool does.
The four steps from script to footage
Your script is cut into scenes
Each scene covers a sentence or two, and a sentence over about 19 words is cut, at a comma when there is one. A section title on its own line (2 to 10 words, no full stop) gets its own short scene.
A director reads the story first
For a longer video the director first writes a short story brief covering the look, the time of day and the main setting, so shots written in separate batches still belong to one video.
One shot is written for each scene
The director writes a short paragraph per scene: the shot type and camera movement, the subject and what they are doing, the setting with a few concrete details, and the light. Every shot has to illustrate what its own narration line says, not just its general topic.
Each shot becomes a clip
A video model generates a clip from each shot. The clip is cut to the length of its narration, captions are added, and the clips are joined into one video.
Generated, not searched
Some tools match a script to stock clips by keyword. Lantaw generates new footage for each line, so it can show a specific moment of your story. It also means the footage is made, not filmed: it illustrates a line and is not a record of a real event.
What the director is told to do
These are the rules written into every shot. They are instructions to the writer and the video model, not guarantees.
- Show what the line says. A shot has to match its own narration, and a line about conflict is shown as conflict, not as friendliness.
- Match the audience country. People look local to the country you pick, with local clothing, architecture and surroundings as the cues. Flags are never used as a cue.
- Keep the period right. When the narration names a year or an era, every shot uses the vehicles, clothing, buildings and tools of that time and nothing modern. With no era, the shot is set in the present.
- Keep one look. Colour and light stay the same across the video, and the time of day and weather change only when the narration moves in time or to a new place. A line with no moment of its own, such as a fact or a tip, gets plain overcast daylight.
- Give each shot a job. A new section gets an establishing shot, a fact gets one concrete detail, a feeling gets a human moment, a question or a “but” gets suspense, and a payoff gets a reveal. Neighbouring shots vary the subject and the camera movement.
- Avoid readable text. AI video draws garbled lettering, so shots avoid signs, documents, screens, logos and flags, and packaged objects are described as plain and unmarked. Short titles such as a year, a number or a name can be added by Lantaw itself as on-screen text, spelled correctly.
- Show only anonymous people. Never a real or recognisable public figure. When the narration is in the first person, the narrator is shown from behind, over the shoulder, as a silhouette or as the camera’s own view, never face-on.
- One continuous take per shot. A shot never dissolves or cuts inside itself. To move between two places, the director writes two shots.
- Turn an abstract line into one concrete thing. A number, price or statistic is shown as an object or an action that stands for it, never as digits on screen.
- Keep hands rare. They are the hardest thing for the video model to draw, so no more than one shot in six shows one.
How to write lines that match
Say where and when
The director has only your words to go on. A line that names a place, a season or an era gets a specific shot. A line that says only that things changed gets a general one.
One idea per sentence
A short sentence maps to one shot. A sentence over about 19 words is cut, and with no comma it is cut in the middle, so add a comma or a full stop where you would pause.
Put section titles on their own line
A short line with no full stop, 2 to 10 words long, is treated as a title and gets its own clip. It also gives the story a clear start of a new part.
Prefer what can be seen
Places, objects and actions turn into good shots. Opinions and statistics turn into one stand-in object, so follow them with a line that shows the thing itself.
Describe the event, not a face
Footage never shows a real named person, and a person’s face will not stay the same from clip to clip. Write what happens and where, and use an AI presenter if you want one consistent face.
Where it can go wrong
- A person does not look the same from clip to clip. Each clip is generated on its own, so only an AI presenter keeps one face, and only on the presenter clips.
- Hands and faces can still warp. The rules keep hands rare, but the video model can draw one badly.
- The rules are not a guarantee. A clip can still show lettering or an invented symbol, because the video model does the drawing.
- The footage illustrates and does not check your facts. Dates, names and claims are yours to verify.
- You cannot pick the footage for a line yourself. You shape the result through the script, the audience country and the presenter and picture shares. Edit clips on a finished video lets you retake one b-roll, picture or presenter clip with a note saying what to change, and pay for that clip only; the whole edit is refunded if the video is not rebuilt.
Clips or still pictures
Some scenes can be an AI-made still picture with a slow zoom instead of a generated clip. A picture is written from its own narration line under similar no-text rules, and costs a little less than a clip. The Create form can fill up to half of a video with pictures. Use them for lines that suit a still and keep motion for the rest.
AI b-roll matching FAQ
How does Lantaw match b-roll to my script?
It cuts the script into short scenes, a scene director writes one shot for each scene, and a video model generates a clip from each shot. The clip is cut to the length of its narration.
Does Lantaw use stock footage?
No. Every b-roll clip is generated from a shot written for its line, and pictures are AI-made images.
Can I choose the footage for each line?
Not directly. You control the result through the script, the audience country, and the share of presenter and pictures. With Edit clips you can retake one b-roll, picture or presenter clip of a finished video, writing a note that says what to change, and you pay for that clip only.
Will the same person appear in every clip?
No. Each clip is generated separately, so a person will not look the same across clips. The look of the video (colour, light and setting) is kept consistent. An AI presenter is the way to keep one face.
Can the footage show a real place or a real person?
It shows a place of the right kind, such as a harbour or a market, in the style of the country you choose. It does not depict a specific named building or institution, and it never shows a real public figure.
What happens to a line that is a number or an idea?
The director picks one concrete object or action that stands for it and never writes digits or prices into the shot. With on-screen text on, a short title such as a year can be added by Lantaw itself.
Try a script and see the shots
Write a short script and pick the audience country, and Lantaw builds the narration, the b-roll and the captions.
See plans and pricing About the AI faceless video generator How to make faceless YouTube videos with AI How much a faceless YouTube video costs