Seedance Prompt From a Video: A Grammar That Renders

Seedance-family models respond to prompts the way a cinematographer responds to a brief: state the subject, state what the camera does, state what changes. Everything else — mood words, genre labels, adjectives about quality — is decoration the model was never trained to prioritise. Here is the grammar, and how a video reference supplies almost all of it.

What Seedance-family prompts reward

ByteDance’s video models are captioned on production-style descriptions rather than on poetic prose, so three properties consistently improve renders: observable statements (what a camera operator would write, not what a viewer would feel), single-movement shots, and explicit end states. It is also strong on camera physics — the parallax change that distinguishes a dolly from a zoom is rendered as a real difference, provided you asked for it.

SocialToPrompt lists Seedance among the models its extracted prompts are written to serve, which is why the drafting step below takes two minutes rather than twenty.

The three-part sentence

Build each shot as one sentence with three ordered clauses:

[subject + framing]  +  [camera behaviour]  +  [what changes / end state]

Close-up of a glass bottle on wet slate, camera low and static  →  slow push-in from chest height
   →  condensation fills frame as the label comes into focus.

• Clause one is nouns and geometry. What is in frame and how much of it.
• Clause two is exactly one verb. Push in, pull back, track, pan, tilt, crane, orbit, handheld, locked-off. If you need “and”, you need two shots.
• Clause three is the delta. Models handle “state A becomes state B” far better than an instruction to be interesting.

Light, grade and material ride on the end as short trailing clauses, because they qualify the shot rather than driving it.

Describing shots it will not infer

Four things a Seedance-class model will not guess, and where renders usually fail:

• Stability. If you do not say “locked-off, tripod-stable”, it adds motion, because most training footage has some.
• Contact physics. “Liquid pours and splashes on impact” beats “pours elegantly”.
• Camera height. The single most under-described property; a 20cm difference changes the whole shot.
• What the frame ends on. Without an end state, the last second becomes a random walk.

The vocabulary for the movement clauses is the same as any other model’s — see extracting camera movement from a video, which covers the dolly-versus-zoom distinction this family renders most visibly.

Multi-shot prompting

Where this family earns its reputation is sequence coherence: several shots that share a subject and a look. Two rules make it work. Repeat the subject noun verbatim in every shot description — a model has no memory between generations, so “the bottle” becoming “a glass container” is how a sequence stops matching. Declare the constants once per shot: light source, palette, format, aspect ratio.

SHOT 1 (9:16, 3s): macro of condensation on a dark glass bottle, locked-off, hard rim light right.
SHOT 2 (9:16, 3s): same bottle on wet slate, slow push-in, identical hard rim light right.
SHOT 3 (9:16, 2s): same bottle, hand lifts it out of frame left, camera stays locked-off.

That shape is a shot list, and the extraction workflow that builds one from a longer video is in extracting shot lists and prompts.

Extraction as the drafting step

Write the grammar by hand once, then stop: extract the reference clip and reformat its fields into the three-part shape. Paste a link — 27 platforms including TikTok, Reels and YouTube — or upload MP4, MOV or WebM, and SocialToPrompt returns the subject, camera, lighting, pacing and style lines its analysis read off the frames, in 20–40 seconds per the site, in around 20 output languages. Cost, stated plainly: 1 credit to extract, 3 to generate the prompt, so a full run on a clip up to 20 seconds is 4 credits, and new accounts get 3 credits valid for 24 hours.

Then do the two edits only you can do: confirm the camera verb against the source, and replace any inherited mood word with the physical clause above. The step-by-step conversion is in converting a video to a Seedance prompt, and the platform-specific version for short-form is TikTok to Seedance.

Frequently asked questions

How do I write a Seedance prompt from an existing video?

Extract the clip's visual fields — subject, camera behaviour, lighting, pacing, style — then reformat them as one sentence per shot in subject, camera, change order. Add the physical details the extractor cannot know: camera height, stability, and what the frame ends on.

What structure works best with Seedance-family models?

Three ordered clauses: subject plus framing, exactly one camera verb, then what changes and how the shot ends. Light, palette and material go last as short qualifying clauses. Multiple camera moves in one prompt is the most common cause of an unstable render.

How do I keep several generated shots looking the same?

Repeat the subject noun verbatim in every shot prompt and restate the constants — light source, palette, aspect ratio, format — each time. Models have no memory between generations, so any drift in your wording becomes drift in the footage.

Does a video extractor replace writing Seedance prompts?

It replaces the drafting, not the editing. Extraction returns the observable fields quickly, but it will understate or invent camera movement on stabilised or tripod footage and summarise grades as mood words, so those lines need a human check.