AI clips

An AI clip is a few seconds of generated film, with sound, that stands in for a scene the website cannot show — a person, a place, a feeling, a before. It sits in the film like any other scene: under the narration and the captions, cut on the same rhythm.

Where clips come from today
Your first clip scene
Choosing a model per scene
Length and resolution
Sound
The same face in every clip
Site films keep the site
Writing a clip prompt
Keeping the cost in hand
When a clip fails

Where clips come from today

Be precise about this, because it is the most common surprise: pasting a link does not produce clips. The automatic planner behind create_video films the site and draws motion-graphic cards; for a beat no page can show it may add an AI still (about two per film at most), and it never adds a clip. A clip is always something you ask for, in one of four ways:

Where	How	Model
Web editor	A scene’s Video clip button: describe the shot, and the clip replaces that scene’s picture	house clip
`create_animation`	An `image` scene with `data.generate` (the shot) and `data.clip: true`	house clip, or `data.model`
`update_storyboard`	`set_visual` a scene to `image` with the same `data`, or `add_scene` with it	house clip, or `data.model`
`generate_asset`	`kind: "video"` makes one clip straight into your media bank, with no film	house clip

A clip in your bank goes into any film at no further charge: put its asset id in an image scene as data.asset_id. A presenter — a person talking in the corner, over the scene rather than instead of it — is also a clip, set with set_inset; see A presenter in the corner.

AI clips are a paid-plan feature. On a free plan a clip scene is drawn as text and the render says so.

Your first clip scene

Over MCP, turn one scene of an existing film into a clip:

{ "name": "update_storyboard",
  "arguments": {
    "video_id": "YOUR_VIDEO_ID",
    "operations": [
      { "op": "set_visual", "scene_id": "YOUR_SCENE_ID", "visual": "image",
        "data": {
          "generate": "A founder at a kitchen table at dawn, laptop open, coffee steaming. Slow push-in. Warm window light.",
          "clip": true
        } }
    ] } }

Then rerender_video. The clip is generated during the render, charged when it is made, and kept on the scene: the next re-render reuses it at no charge. Changing the generate text makes a new one.

In a storyboard you write from scratch, the scene looks the same:

{ "visual": "image",
  "heading": "The old way",
  "narration": "Every morning started the same way: forty tabs and no answer.",
  "data": { "generate": "A tired designer rubbing his eyes in front of a wall of browser tabs, blue screen light, slow dolly out.",
            "clip": true } }

Choosing a model per scene

Leave out model and the scene gets the house clip: Veo 3.1 Fast, 8 seconds at 720p, with sound, for a flat 311 credits. Name a model and the scene is made by that model and charged its price, per second at the resolution it makes:

"data": { "generate": "…", "clip": true, "model": "seedance-2.0-mini" }

The choice is per scene, not per account: a draft-quality beat can use seedance-2.0-mini and the hero shot seedance-2.0. Veo 3.1 Standard (veo-3.1) is used only when a scene names it.

Which models exist, and which you can use now: list_models (free).
What each costs: the generated table in Credits & pricing.
How the families differ: AI models.

Length and resolution

You do not set a clip’s length or resolution in a film — they are planned from the shot, within what the model makes:

Length is how long the scene will be on screen: its narration at about 110 words a minute plus a breath, or its duration_seconds if it is silent. A model with a fixed menu of lengths gets the shortest one that covers the scene. The house clip is always 8 seconds.
Resolution follows how big the clip is drawn: at least 720p (768p on MiniMax) for a clip that fills the frame, the model’s cheapest tier for a presenter in a corner.
A scene longer than the model’s maximum holds the clip’s last frame for the rest of the scene, and the render says so. Shorten the narration, or pick a model with longer clips.
Nothing is silently changed at quote time. quote_cost refuses a length or resolution the model does not offer, with the ones it does in allowed. During a render, if a model refuses a length it advertises, it is asked once more at 5 seconds, the length every model has proven, and ai_report records it.

Sound

Every clip is generated with its own sound. Under a narrated scene it plays as a quiet bed beneath the voice; in a scene with empty narration it plays at full volume and carries the scene. A line in quotation marks inside generate is spoken by the person in the shot, and a silent scene is held until their last word, measured from the clip itself, so nobody is cut off mid-sentence.

The same face in every clip

A character is a name and one to three reference images. Add data.character to a clip scene and those images go to the model with the prompt, so the same person appears in every shot, whichever model makes it:

"data": { "generate": "Ada walks into a bright co-working space and waves at the camera.",
          "clip": true, "character": "Ada, the founder" }

Site films keep the site

A clip replaces a scene’s picture. In a film of a website, the site is the subject, so PageToVid keeps at least half of the scenes as real recordings of the site (one in four in a 9:16 film, never fewer than two in a film of four scenes or more). The planner builds to that floor and the storyboard review flags a film that falls below it: add clips in place of drawn cards, not in place of every recording. Films you write with create_animation film no page, so they have no floor — a film can be all clips.

Writing a clip prompt

A clip is one shot, a few seconds long. Prompts that work describe one:

Do	Why
One subject, one action. “A barista slides a cup across the counter.”	Several actions in eight seconds come out as a blur of half-actions.
Name the camera move. Slow push-in, dolly out, handheld follow, static wide.	Without one, most models drift aimlessly.
Name the light and the mood. Warm morning window light; cold office fluorescents.	It sets the look faster than any adjective.
Say where we are. A small bakery at opening time; a crowded train.	A subject with no place floats.
Put speech in quotation marks. He says: “Two minutes, that’s all it took.”	A quoted line is spoken; keep it short, one sentence.
No on-screen text, logos, UI or brand names.	Generators draw text badly, and a fake logo or a mangled interface costs trust. Show the real product with a recording; let the clip carry the feeling.

For the same person across scenes use a character rather than re-describing them: a description drifts from shot to shot, reference images do not.

Keeping the cost in hand

Price first: quote_cost prices one clip (kind, model, duration_s, resolution), or with a video_id the next re-render of a whole film — render plus every visual not made yet. Free.
Cap it: create_animation’s ai_budget sets max_clips and max_images for the render, and ai_plan in the response shows the bill before anything is made.
Draft cheap, finish sharp: iterate on seedance-2.0-mini, then switch the one hero scene to a bigger model.
Reuse: a clip already made is kept on its scene and in your bank; re-rendering, moving a presenter or using a bank clip via data.asset_id costs nothing more.

When a clip fails

The render never stops for a failed clip. The clip is refunded, the scene falls back to its page still and its text, and get_video says so in warnings and ai_report, with the model’s own reason: a refusal on content grounds means change the words, a timeout means try again.