Blog

Guide

Seedance 2.5 Text-to-Video Prompt Recipes You Can Copy and Paste

Single-shot 30s storyboard structure, camera and pacing wording, native audio cues, and negative prompts — with notes on what real outputs look like

Nov 18, 2026 12 min read
Cover art for a Seedance 2.5 prompt guide: a flat loom-diagram weave of broad espresso, oat and brass-toned textile bands.
On this page
One 30s single-shot prompt for Seedance 2.5 — camera, pacing, lighting and native audio cues in a copy-paste recipe.

What makes a text-to-video prompt stable

Most prompt failures on Seedance 2.5 don't come from bad ideas — they come from ambiguous instructions that the model resolves differently on every run. A prompt that says "a cinematic chase scene" leaves the model to invent the subject, the camera, the number of cuts, and the pacing. Every generation becomes a coin flip.

Stable prompts share one property: every component is explicit and ordered. Think of the prompt as six stacked layers that all describe one continuous shot:

  • Subject — who or what is on screen, with enough visual detail to lock identity (age, wardrobe, material, color palette).
  • Action beats — what happens, in the order it happens, as a sequence of visible events.
  • Camera — one continuous path and one movement vocabulary, not a list of shots.
  • Pacing — how fast beats unfold relative to each other.
  • Audio — diegetic sound tied to the visible action.
  • Negatives — what must not appear.

The critical constraint is that Seedance 2.5 text-to-video is strongest when the prompt describes a single unbroken camera move. The moment your prompt implies multiple locations or cuts, the model has to invent transitions, and transitions are where drift, warped motion, and style breaks show up first. That's why the recipe skeleton below puts everything on one timeline instead of writing scene-by-scene instructions.

A second rule: describe, don't label. "Slow dolly-in" is a label. "The camera pushes forward at walking speed, closing from a wide frame to a medium close-up by the final beat" is a description. Labels get reinterpreted; descriptions get followed.

The 30-second single-shot recipe template

Here's the base structure. Fill in the brackets and keep the order — the ordering is what keeps the camera continuous.

Single continuous shot, 30 seconds, no cuts.

Subject: [one person / object / creature], [age or scale], [wardrobe or material], [distinguishing detail].

Setting: [location], [time of day], [light quality], [weather or atmosphere].

Beats (in order, one continuous camera):
1. 0-8s: [opening action that establishes the subject and space]
2. 8-15s: [the subject begins the main action; camera starts its move]
3. 15-23s: [the peak moment — the clearest, most readable action]
4. 23-30s: [settling action; camera comes to rest or holds]

Camera: [one movement path], starting [framing] and ending [framing].

Pacing: [overall rhythm], with the fastest motion around beat 3.

Audio: [ambience bed], plus [sound tied to beat 2], [sound tied to beat 3].

Avoid: [short negative list].

Why four beats

Four beats map cleanly onto 30 seconds and give the model a readable rhythm. Fewer beats and the model pads with invented filler; more beats and it rushes, which reads as speeding up or cutting. Each beat should describe something visible changing — a hand moving, a door opening, dust settling — not an internal state.

Keep the camera sentence singular

Write exactly one camera path and reference it across all beats. "The camera starts wide and slowly arcs left around the subject, ending behind their shoulder" is one path. If you write a second camera sentence, treat it as a second shot and expect a cut.

Beat 3 is your anchor

If you only get one beat right, make it beat 3. Put the most visually legible action there — the thing a viewer would describe if they saw only a still frame. The other beats exist to lead into and out of it.

Camera and pacing wording that actually holds

Vague camera language is the top cause of unintended cuts. Contrast these pairs:

  • Vague: "dynamic camera work" → Concrete: "the camera tracks right at a steady walking pace, keeping the subject centered."
  • Vague: "cinematic angles" → Concrete: "starts at a low wide angle, rises slowly to eye level by the final beat."
  • Vague: "fast editing" → Concrete: "motion within the frame accelerates; the camera itself stays continuous."

A small vocabulary that behaves well

These movement descriptions map to physically plausible camera paths and tend to hold across a full 30 seconds:

  • Push in / dolly forward — specify start and end framing (wide → medium close-up).
  • Pull back — specify what gets revealed as the frame widens.
  • Arc / orbit — specify direction and degrees, e.g. "a slow quarter arc to the left."
  • Track — specify direction and that the subject stays framed.
  • Crane up / lower — specify the height change in plain terms.
  • Static with internal motion — the camera holds while the subject moves; good for product and portrait shots.

Pacing as relative speed, not adjectives

Instead of "fast-paced," describe how beat timing relates:

Pacing: the first two beats unfold slowly and deliberately;
motion builds through beat 3; the final beat settles and holds still.

This gives the model a curve rather than a constant. Curves read as intentional; constants read as either sluggish or frantic.

One movement per shot

Resist combining movements — "push in while orbiting and tilting up" asks for a rig path most real cameras can't perform, and the model resolves the contradiction by cutting. Pick one primary movement and, if needed, one small secondary motion like a slight rise.

Writing native audio cues into the prompt

Seedance 2.5 generates audio alongside video, but audio only stays coherent when it is written as a consequence of visible action. Detached sound descriptions — "epic music, cinematic sound design" — tend to produce generic beds that don't sync with anything.

Structure the audio line in three parts

Audio: [continuous ambience bed] + [sound caused by beat 2 action] + [sound caused by beat 3 action]. No music.

The ambience bed establishes the space (room tone, wind, distant traffic). The action sounds anchor sync. Naming whether you want music — and explicitly excluding it when you don't — prevents the model from adding a score that fights your dialogue or voiceover later.

Tie every sound to something on screen

  • Weak: "footsteps and a door." → Strong: "the sound of boots on wet concrete as the subject crosses frame in beat 2, then a heavy door latch clicking as they push through in beat 3."
  • Weak: "crowd noise." → Strong: "low, indistinct crowd murmur behind the subject, rising slightly as the camera pushes in."

Useful audio phrasings

  • "room tone only, no music"
  • "ambience continues uninterrupted across the whole shot"
  • "sound rises and falls with the visible motion"
  • "no voiceover, no dialogue"

If you plan to add voiceover in post, say so in the prompt ("no dialogue") so the model doesn't generate competing speech. Audio drift is usually a sign that the audio line described sounds the picture doesn't show.

Negative prompt checklist

Negatives work best as a short, specific list aimed at the failure modes you actually see. Long generic lists dilute attention and can suppress legitimate elements.

The reusable base list

Avoid: cuts, scene changes, multiple shots, style drift,
warped hands or faces, extra limbs, text overlays, watermarks,
logos, subtitles, flickering, sudden zoom, speed ramps.

What each cluster prevents

  • cuts, scene changes, multiple shots — the single biggest stabilizer for long prompts.
  • style drift — keeps lighting and color grade consistent from beat 1 to beat 4. Add a positive style anchor too ("consistent warm interior lighting throughout").
  • warped hands or faces, extra limbs — anatomy stability, especially during fast motion in beat 3.
  • text overlays, watermarks, logos, subtitles — models sometimes invent signage or captions; explicit exclusion saves cleanup.
  • flickering, sudden zoom, speed ramps — pacing artifacts.

Keep it under roughly a dozen items

If you're past about twelve negatives, you're probably compensating for a vague positive prompt. Fix the description first, then trim the list. Remove any negative that hasn't appeared in your recent outputs — stale negatives can block details you want.

Match negatives to the shot

Product shots need "no hands entering frame" and "no reflections of a camera crew." Portrait shots need "no background extras." Add shot-specific negatives, keep the base list, and re-test.

A copy-paste recipe library

These are working baselines. Adapt the bracketed details; keep the structure intact.

Recipe 1 — product hero, static camera

Single continuous shot, 30 seconds, no cuts.
Subject: a matte ceramic pour-over coffee dripper, [color], resting on a light oak counter.
Setting: minimal kitchen, soft morning window light from the left, calm atmosphere.
Beats:
1. 0-8s: the dripper sits still; steam rises gently from the carafe below.
2. 8-15s: a hand enters from the right and slowly lifts the dripper.
3. 15-23s: coffee pours in a steady stream into the carafe, catching the light.
4. 23-30s: the pour stops; the hand sets the dripper down and exits frame.
Camera: static, locked off, slight rise from counter level to a gentle three-quarter angle.
Pacing: slow and deliberate throughout; motion peaks in beat 3.
Audio: quiet room tone, soft pour sound in beat 3, ceramic click in beat 4. No music, no dialogue.
Avoid: cuts, text overlays, logos, warped hands, flickering.

Recipe 2 — character walk, tracking camera

Single continuous shot, 30 seconds, no cuts.
Subject: a [age] person in a [color] raincoat, hood down, carrying a canvas bag.
Setting: narrow city street after rain, overcast dusk, wet pavement reflections.
Beats:
1. 0-8s: the subject walks toward the camera at an easy pace.
2. 8-15s: the subject turns and begins walking alongside the camera.
3. 15-23s: they step over a puddle, bag swinging, coat catching the wind.
4. 23-30s: they slow and stop, looking off frame.
Camera: tracks right at walking pace alongside the subject, keeping them framed at medium distance.
Pacing: steady rhythm, slightly faster through beat 3, settling in beat 4.
Audio: light rain ambience, footsteps on wet concrete in beats 1-3, distant traffic. No music.
Avoid: cuts, style drift, warped face, extra limbs, subtitles, speed ramps.

Recipe 3 — environment reveal, pull-back camera

Single continuous shot, 30 seconds, no cuts.
Subject: a lone wooden cabin on a snowy ridge.
Setting: alpine dawn, low golden light, thin mist in the valley.
Beats:
1. 0-8s: close on frost on the cabin window; faint warm light inside.
2. 8-15s: the camera begins pulling back, revealing the cabin door and porch.
3. 15-23s: the pull-back continues, revealing the ridge and the mist-filled valley.
4. 23-30s: the camera settles on a wide landscape; smoke rises from the chimney.
Camera: slow continuous pull-back from close-up to extreme wide.
Pacing: unhurried, constant reveal, no acceleration.
Audio: wind and faint snow crunch, distant birds. No music, no dialogue.
Avoid: cuts, zooming, text overlays, flickering, style drift.

For a full list of supported parameters and current model capabilities, see the Seedance 2.5 text-to-video model page.

Testing outputs and iterating one variable at a time

Prompting is empirical. The loop below turns random re-rolls into a repeatable process.

1. Freeze the seed where possible

If the API exposes a seed, keep it fixed while changing prompt text. You're testing whether a wording change produces a behavior change, not whether the model got lucky.

2. Change exactly one variable per run

Run A is your baseline. Run B changes only the camera sentence. Run C changes only the audio line. If you change three things and the output improves, you've learned nothing reproducible.

3. Watch for the same three signals

  • Camera continuity — did the shot stay unbroken?
  • Beat timing — did action land roughly where the timestamps said?
  • Audio sync — did sounds match visible events?

Score each pass/fail in your log; patterns emerge fast.

4. Keep a prompt log

Store, per run: the full prompt text, the seed, the parameter values, and a one-line note on the result. When a generation works, you want to reproduce it exactly, not approximate it from memory.

5. Batch comparable runs

When tuning the camera sentence, generate several variants of the same prompt back to back and compare. Side-by-side comparison beats sequential memory.

The rule of thumb

If two consecutive runs with the same prompt differ wildly, the prompt is under-specified — go back and add detail to the subject and camera layers before tuning anything else.

Common failure modes and how to fix them

Symptom: the shot cuts or changes location mid-clip

Cause: multiple camera sentences or implied scene shifts. Fix: collapse to one camera path, add "single continuous shot, no cuts" at the top, and put "cuts, scene changes, multiple shots" in negatives.

Symptom: style drifts — lighting or color changes across the 30 seconds

Cause: no positive style anchor. Fix: add a consistency clause ("consistent warm interior lighting throughout") and keep "style drift" in negatives.

Symptom: the model rushes or pads the middle

Cause: beat timing is vague or there are too few/many beats. Fix: use four beats with timestamps and add a pacing curve sentence ("motion builds through beat 3, settles in beat 4").

Symptom: audio is generic or out of sync

Cause: sound described without visible causes. Fix: tie each non-ambient sound to a specific beat's action; explicitly say "no music" if you don't want a score.

Symptom: warped hands or faces during fast motion

Cause: too much simultaneous movement in one beat. Fix: slow beat 3 slightly, reduce simultaneous actions, and add anatomy negatives ("warped hands or faces, extra limbs").

Symptom: unexpected text, signage, or captions appear

Cause: the model invents labels in busy scenes. Fix: add "text overlays, watermarks, logos, subtitles" to negatives and simplify background detail.

Symptom: motion looks like a speed ramp

Cause: pacing described as "fast" without a curve, or conflicting speed cues. Fix: describe relative speed per beat and add "no speed ramps" to negatives.

Start from a known-good recipe, change one thing, and log the result — that's the fastest path from a clever prompt to a reliable pipeline.

Related articles

View all articles →

Start building with persistent AI artifacts

Get your API key and call image, video and audio models through one task-based API.

Read the API docs