Seedance 2.5 Prompts: Formula and 15 Examples With References

Seedance 2.5 prompts work best written as a short director's script: who is in frame, what happens second by second, how the camera moves and which file does what — @Image1 for the face, @Video1 for motion, @Audio1 for music. In the IIshki studio, Seedance 2.5 builds clips up to 30 seconds in 480p, 720p or 1080p, either from a first frame or from a set of references: up to 30 photos, 10 videos and 10 audio tracks. Below are the prompt formula, 15 ready-made examples by task, and the mistakes that make the model mix up references or drift halfway through a clip.

For short clips up to 15 seconds and cheap drafts, the same techniques work in Seedance 2.0: it takes up to 9 photos, 3 videos and 3 audio tracks and has a budget Mini plan. How both models stack up against Kling is covered in the Seedance 2.0 vs Kling 3.0 comparison.

The Seedance 2.5 prompt formula

Long clips rarely break on the image itself — they break on transitions. The model doesn't know what should happen at second 12, so it repeats one action. That's why a 2.5 prompt is written in blocks, not as one paragraph. A working order:

1. Style — genre, light, film or digital look, color palette. One line that sets the tone for the whole clip.
2. References — the role of each file: "@Image1 — her face and hair, @Image2 — her coat, @Video1 — the camera move".
3. Location and starting position — where we are and how the shot opens.
4. Shot breakdown — timecoded segments: "0–6 s … 6–14 s …". One action and one camera move per segment.
5. Sound — what plays and what shouldn't: "music @Audio1, no dialogue".

A template you can drop any scene into:

Style: [genre, light, palette]. References: @Image1 — [role], @Video1 — [role]. Location: [where, time of day]. 0–8 s: [action], camera [move]. 8–16 s: [action], camera [move]. Sound: [what plays], no [what to avoid].

A 5–8 second scene doesn't need blocks — three to five sentences will do. A breakdown pays off from 12–15 seconds and is a must for 20–30.

Which mode and model to use for the task

If you want to replicate someone else's scene end to end, there's a step-by-step guide on how to clone a video with AI.

How to reference files: @Image1, @Video1, @Audio1

Numbers follow upload order and are counted separately per type: the first photo is @Image1, the first video @Video1, the first track @Audio1. The model can't guess why you uploaded a file, so give each one a role:

• Face and look — "keep the face, hair and build from @Image1 in every shot".
• Clothes and props — "the hero wears the jacket from @Image2 and holds the bag from @Image3".
• Location — "the action takes place in the interior from @Image4".
• Motion — "the camera repeats the orbit from @Video1", "cut rhythm as in @Video1".
• Sound — "the clip plays to the music in @Audio1, movements on the beat".

Two or three photos of one hero from different angles beat a single shot: the face drifts less when the head turns. But 30 files is a ceiling, not a target. Five well-chosen references give a more coherent clip than twenty where three similar faces and two different interiors compete.

15 ready-to-use Seedance 2.5 prompts

Copy them and swap the hero and location for your own. The @-numbers must match your upload order.

Hero and story

Style: warm evening light, film grain. @Image1 — the hero's face and hair. 0–7 s: he walks out of the building, camera follows from behind. 7–15 s: he gets into a car, camera slowly pulls back. 15–20 s: the car drives off down a night street, wide shot from above. No dialogue.

@Image1 and @Image2 show the same woman; keep her face and proportions. She walks through an autumn park, picks up a leaf and smiles at the camera. The camera moves ahead of her on a gimbal. 10 seconds, soft backlight.

Style: handheld documentary. @Image1 — a craftsman in the workshop from @Image2. 0–6 s: close-up of hands sanding a board. 6–12 s: medium shot, he blows away the dust. 12–18 s: he lifts the finished shelf toward the light. Sound: sandpaper, quiet workshop, no music.

Products and ads

@Image1 — a perfume bottle; keep the shape and label. The bottle stands on wet black stone, water drops run down the glass, the camera slowly orbits. Warm backlight, 8 seconds.

Style: clean studio commercial. @Image1 — a sneaker. 0–5 s: the sneaker spins on a white background. 5–10 s: it breaks into parts and reassembles. 10–15 s: final frame, sneaker centered, shadow below. Music @Audio1.

@Image1 — a young woman, @Image2 — a jar of cream. She stands by a window, applies the cream to her hand and looks into the camera. Natural daylight, shot like on a phone, no studio polish. 12 seconds.

Motion from a sample (@Video1)

Repeat the camera move from @Video1 for the woman in @Image1 on a city rooftop at dusk. Warm sunset light, 15 seconds.

@Video1 — pacing and cut rhythm. Build a clip in the same rhythm: @Image1 runs along the embankment, @Image2 — a view of the bridge, @Image3 — a final shot by the water. Music @Audio1.

The character from @Image1 performs the dance from @Video1 in an empty gym. Static camera, wide shot, 10 seconds.

Editing and extending a finished clip

For these prompts, set "Video task" to "Edit" or "Extend" and upload the source as @Video1.

Edit: in @Video1, turn daylight into night, switch on neon signs, add reflections on wet asphalt. Same motion and hero.

Edit: in @Video1, dress the hero in the suit from @Image1, everything else unchanged.

Extend: after the last frame of @Video1, the hero turns, walks to the door and leaves; the camera stays put. 8 seconds.

Mood and landscapes

Style: epic landscape, 21:9 widescreen. 0–10 s: fog over a mountain valley, the camera slowly rises. 10–20 s: the fog clears to reveal a lake. 20–30 s: a flyover above the water at sunset. Sound: wind, no music.

A quiet old-town street at night, light snow, streetlamps, the camera glides past the houses at eye level. 15 seconds, no people in frame.

Style: cartoon with soft shapes. @Image1 — the character. He wakes up, stretches and looks out the window at a sunny morning. 10 seconds, music @Audio1.

How to run a prompt in the studio: 4 steps

1. Pick the input mode

Open Seedance 2.5 in the studio. Choose Frames if you have a ready first frame, References if the scene is built from several photos, videos and tracks. No photo of the hero? Generate one in Nano Banana Pro, and the music in ElevenLabs Music.

2. Upload files in the right order

Upload order is what sets @Image1, @Image2. Hero first, then clothes and props, then the location. Videos and audio are numbered separately.

3. Paste the prompt and set parameters

Duration from 4 to 30 seconds, aspect ratio from 9:16 to 21:9 or adaptive, quality 480p, 720p or 1080p. Test in 480p: debugging the prompt costs less, then rerender the final in 1080p.

4. Change one thing at a time

If the result is off, change one variable per run: a reference's role, one segment of the breakdown, or the camera move. Change everything at once and you won't know what helped. The run cost is shown on the model card before you start.

Common Seedance prompt mistakes

What's next

Start with one prompt from the list at 8–10 seconds in 480p, get a stable hero, and only then stretch the scene to 20–30 seconds. Need prompts for another model? See 20 examples for Kling 3.0, and for picking a video model, the roundup of the best AI video generators. Multi-step scenes are easy to assemble in Canvas, where the hero photo, the music and Seedance connect as nodes in one chain.