Seedance 2.5 30-second video prompts: stages and timing

A five-second clip is a description. A thirty-second clip is a script. That shift is the whole reason people who write good short prompts suddenly get mush when they ask for length: they keep writing a richer scene when what the model now needs is a sequence of checkpoints. Seedance 2.5 is available in the OmniArt video workspace, and it generates 4 to 30 seconds in a single pass — plus 30 to 180 seconds in Long Video mode — which means the structure of your Seedance 2.5 30-second video prompts matters more than the adjectives in them.

The approach below comes from how the model actually reads a long prompt. You divide the story into consecutive stages, give each stage exactly one primary state change, and write an end state that a viewer could point at on screen. Only then do you layer timestamps, and only where you need real precision.

This is not a return to numbered shot lists. A shot list tells the model how to operate the camera; a stage tells the model what must be true when the stage ends. If you want the argument against over-specifying short clips, stop writing shot lists still holds. Stages are state checkpoints, not camera settings.

Long videos fail differently from short ones

A four-second clip rarely has time to contradict itself. Thirty seconds does, and the failures cluster in four places.

• Character identity. The person at second 2 and the person at second 24 are cousins rather than the same individual.
• Clothing and props. A jacket shifts shade, a ribbon changes color, a tool that was on the bench is suddenly in a hand — or vice versa.
• Prop ownership. The model loses track of who is holding what across a handoff, which is the single most common break in any two-person sequence.
• Spatial direction. The workbench faces the other way after a cut, or a character who exited left re-enters from the same side.

None of these are fixed by more description. They are fixed by telling the model, at each boundary, what state it is inheriting. Seedance 2.5 holds character features, clothing detail and scene architecture noticeably better across a long generation than earlier versions, but it still needs the checkpoints to hold them against.

Script stages, not scenes

The long-video structure is a small set of labeled blocks. Copy this and fill it in.

[Generation Goal]
Generate a <video type>. The central subject is <subject>, and the primary event is <story summary>.

[Stage 1]
Initial state: <initial state of characters, props, and scene>.
Primary event: <one primary action or event>.
End state: <character positions, prop ownership, or visible scene state>.

[Stage 2]
Continue from the previous stage: <state that must remain unchanged>.
Primary event: <one primary action or event>.
End state: <observable state>.

[Stage 3]
Primary event: <closing event>.
End state: <final visible state>.

[Maintain Consistency]
Keep <character identity, number of characters, clothing, prop ownership, spatial direction, and audio relationships> consistent.

Three stages is a good default for a 30-second generation. Here it is filled in for a repair-café sequence — two people, one object that changes hands twice, which is exactly the shape that drifts if you write it as one paragraph.

[Generation Goal]
Generate a documentary-style instructional video. The central subjects are <Volunteer> and <Owner>, and the primary event is repairing a desk lamp with a frayed cord and returning it working.

[Stage 1]
Initial state: <Volunteer> stands behind a wooden repair bench in a bright community hall. The desk lamp sits unlit on the bench, and a small metal tray and screwdriver lie to the right of it.
Primary event: <Volunteer> unscrews the lamp base and lifts it away, exposing the wiring.
End state: the lamp base is open on the bench, three screws are in the metal tray, and the screwdriver is back on the right side of the bench.

[Stage 2]
Continue from the previous stage: both characters keep the same identities and clothing, the lamp base stays open on the bench, and the screws stay in the tray.
Primary event: <Volunteer> cuts away the frayed cord and fits a new braided cord into the base while <Owner> holds the lamp head steady.
End state: the frayed cord lies coiled at the left edge of the bench, the new braided cord runs from the base to the floor socket, and <Owner> still holds the lamp head with both hands.

[Stage 3]
Primary event: <Volunteer> closes the base, plugs the cord in, and <Owner> presses the switch.
End state: the lamp stands upright and lit in the center of the bench, and both characters look at the lit lamp from behind the bench.

[Maintain Consistency]
Keep <Volunteer> and <Owner>'s identities, clothing, the bench orientation, the tray position, and lamp ownership at each handoff consistent. Only two people are in frame throughout. Pure video, no subtitles, no background music.

One state change per stage

The discipline that makes this work is boring and absolute: one primary event per stage. If a stage contains "unscrews the base, replaces the cord, and tests the lamp," the model has to allocate its own time budget across three beats and will usually rush two of them or drop one. Split it.

End states must be visible

An end state is a frame you could screenshot, not a feeling. "The repair is finished" gives the model nothing to converge on. "The lamp stands upright and lit in the center of the bench" gives it a target, and it also gives the next stage its inherited state for free. Write end states in terms of positions, ownership, and visible scene state — where things are, who is holding them, what is on or off.

Timestamps, where precision actually matters

Seedance 2.5 supports plot and shot interval control by timestamp natively, accurate to the second. In the launch guidance you write an interval marker such as [00.0-05] directly in front of the shot, line, or action, then script in chronological order.

Use stages by default. Reach for one-second precision when you need to control a handoff, an entrance or exit, a transition, or an explicit beat — a music sync, a product reveal, a cut you have already storyboarded.

The rules are short:

• Consecutive and non-overlapping. Ranges tile the timeline. Gaps and overlaps produce exactly the confusion you would expect.
• A range is a time budget, not an edit point. Actions may land slightly before or after a boundary. If your deliverable needs frame-accurate cuts, that is a post-production job.
• Content density sets the feel. Too little content in a range hands the model freedom and it will invent filler. Too much causes excessive cutting or dropped events.
• Never demand a frequency. "Complete three actions in one second" is not a pacing instruction; it is a request for a mess.

Timestamps and stages compose well: keep the stage blocks, and attach a range to the one or two stages whose pacing you actually care about.

Global settings and the anti-collapse block

The launch guide adds one block above the stages for long prompts: a global settings section that fixes the worldview once so it does not have to be re-derived in every stage. It covers base environment and texture, visual style, camera language, character styling, the performance core, and a list of prohibited items.

That last list is the useful part. Seedance 2.5 improved at honoring negative instructions, so the split between a positive line (frame content, camera movement, action, dialogue, sound effects) and a reverse line (what must not appear) is worth writing per stage as well as globally. "No subtitles, no background music" is now a real instruction rather than a wish, and the recommended phrasing is direct — "pure video, no subtitles, no background music". Add the failure modes specific to your shot: no extra people, no large body movements, no invented dialogue.

Close with a consistency block

The [Maintain Consistency] block is not decoration. Restate, in one place at the end: character identity, the number of characters in frame, clothing, prop ownership, spatial direction, and audio relationships (who speaks, what stays silent). It costs two lines and it is the cheapest defense against the drift patterns listed at the top of this article.

Which mode: single generation, extension, or Long Video

Three ways to get length out of Seedance 2.5, and they are not interchangeable.

Extension is the "segmented director" path: generate ten seconds, look at it, then instruct the next segment. Duration is settable per extension, but aspect ratio is inherited from the input and cannot be changed. Long Video mode skips the stitching entirely for a 30-to-180-second piece — promotional films over landscape, music videos, monologue-driven shorts where an emotional arc needs room.

For an ultra-long prompt the launch guide adds one element on top of everything above: restate the global parameters — target duration and aspect ratio — at the very beginning of the prompt text. The generation is long enough that saying it once in the UI and once in the prompt is the safer habit. The rest of the shape is familiar: global parameters, a creatives description naming which uploaded material is which, a one-sentence overview, the specific plot description as a timeline, and a global supplement at the end.

Output resolution is 480p or 720p, so plan finishing accordingly on long pieces.

Honest limits

• Timestamps allocate time to events. They are not frame-accurate edit points.
• Editing prompts raise the probability that critical events align with the source; they do not guarantee frame-by-frame overlap.
• Multi-reference work is about selecting the correct material for each scene, not making every material appear at once.
• For exact subtitles, formulas, signage, product specs, or frame-level timing, combine prepared reference material, generation, and post-production. Do not expect the prompt to carry it alone.
• Extension boundary frames connect naturally at a visual level rather than pixel-identically, and an extended segment's volume may differ slightly from the source.

Getting started on OmniArt

Take a 30-second idea you have already failed at once. Rewrite it as three stages, give each stage one primary event and an end state you could screenshot, and add the consistency block at the bottom. Generate it on Seedance 2.5 in the video workspace, then look only at the stage boundaries — that is where you will see whether the checkpoints held.

If a boundary breaks, the fix is almost always to split that stage rather than to add description. From there, what shipped in Seedance 2.5 covers the wider feature set, and video editing and extension picks up where a single generation runs out of room.