AI video clips and B-roll

Set clip on an image scene and instead of a still you get a few seconds of generated video: a moving camera, a subject in motion, real depth of field. It is composited like any other visual, under the same captions, with the same transitions.

The clip is generated at the film's aspect ratio, so a 9:16 film gets a vertical clip rather than a letterboxed landscape one. If the scene's narration outlasts the generated footage, the clip holds on its last safe frame rather than cutting to black.

Spend a clip where motion carries meaning, not decoration.

• The opening shot. The first two seconds decide whether anyone watches the rest, and movement buys attention that a static image does not.
• A transformation. Before-and-after, chaos-to-order, a process completing. The change over time *is* the message.
• A human moment. A face reacting reads as a person; a still of a face reads as stock photography.
• The closing beat, if you want the film to feel like it ends rather than stops.

Where a still is the better buy: anything you narrate over for more than four or five seconds, any beat the viewer needs to read, and any beat that is really about your product — because that one should be a real screen recording, not generated at all.

The house clip — Veo 3.1 Fast, eight seconds at 720p, with sound — is 160 credits. A still is 40. The render itself is 40.

So a 40-second film with one opening clip and two supporting stills costs 40 + 160 + 40 + 40 = 280 credits. The same film with five clips costs 840. On the Starter plan's 1,000 credits a month, the first fits three times over; the second fits once.

Name a model instead and the clip is priced from what that model really costs, per second at the resolution used — and since PageToVid asks for a clip as long as the scene is on screen, a longer scene costs more. For an eight-second shot that comes to about 249 credits for Seedance 2.0 Mini or MiniMax H3, about 373 for Seedance 2.0 Fast, 435 for Wan 3.0 Prime, 466 for Seedance 2.0, 718 for Seedance 2.5 and 1,242 for Veo 3.1. Prices follow the providers, so ask your assistant to run list_models or quote_cost first — both are free.

This is why ai_budget matters: set it, and scenes beyond the cap fall back to a still or a card instead of quietly emptying the balance. ai_plan shows you the whole bill before a single generation runs.

Clip descriptions differ from still descriptions in one way: say what moves.

• "Slow push in on a freelance designer at a cluttered desk, one monitor lighting the room."
• "Overhead shot, hands sorting a scattered pile of paper into three neat stacks."
• "Handheld, following a warehouse worker walking down an aisle of shelves."

Name the camera move (push in, pull back, pan, handheld, static), the subject's action, and the light. Keep it to one action — a clip is a few seconds and cannot carry a sequence of events.

Keep the shot simple. Generated video is strongest on a single subject doing one thing and weakest on crowds, text, hands doing fine work, and anything with a brand on it.

Clips are the slowest thing in the pipeline by a wide margin. A film with several of them takes noticeably longer than one without.

Generations do run in parallel rather than one after another, so four clips cost roughly one round of waiting rather than four. But if you are iterating on a script, it is much faster to render with stills, get the words and the timing right, and only then swap the stills for clips on the beats that deserve them. update_storyboard lets you do exactly that without re-authoring the film.

Two clips that both say "a freelance designer" will show two different people. In a 40-second film, that reads as an error.

Create a character first, then reference it by name in each scene. The same face comes back in every clip and every still that names it, which is what makes a multi-shot story about one person possible. This is the single biggest quality difference between a film that uses generated people well and one that does not.