Seedance 2.5 Prompt Guide: References, Roles & 30-Second Scenes

Start With What Already Works

Seedance 2.5 doesn't replace the prompting grammar you learned on 2.0; it extends it. If you haven't internalized the fundamentals, read our Seedance 2.0 prompt guide first, because every technique in it still applies: the layered shot description, precise visual verbs, one strong lighting keyword, real cinematography references, and the warning that vague subjects produce generic output.

The core formula is unchanged:

Subject + Scene/Environment + Action/Motion + Camera Movement + Lighting + Style/Mood

What's new in 2.5 is scale: clips up to a native 30 seconds, up to 50 reference inputs, role tagging, and audio that's generated in the same latent space as the picture. Each of those changes how you should structure a prompt. A quick caveat: 2.5 is still rolling out and some details below come from early API documentation, so treat the specifics as current best knowledge rather than final spec. Our Seedance 2.5 overview tracks what's confirmed.

Structure 30 Seconds as Beats, Not a Description

A 15-second prompt can be one continuous description. A 30-second prompt written the same way gives the model too much room to wander. The fix is time segmentation: break the clip into 3 to 5 explicit beats and describe each one.

0-6s: A woman in a red raincoat stands alone at a rain-soaked bus stop at night, sodium streetlights overhead. Static wide shot. She checks her watch.
6-14s: A vintage car pulls into frame, headlights flaring through the rain. Slow push in as she leans down to the passenger window.
14-24s: Inside the car. Warm dashboard glow. She laughs at something the driver says, tension dissolving. Rack focus from her face to the rearview mirror.
24-30s: Exterior. The car pulls away down the empty street, taillights receding. Crane up to the city skyline. Generate audio: rain, engine, distant jazz.

Three rules for beats:

• Give each beat one job. One location change, one emotional turn, or one camera move. Beats that try to do three things produce mush.
• Make transitions explicit. "Inside the car" tells the model to cut; without it, you get a smeared morph between locations.
• Front-load what matters. The first 20 to 30 words still carry the most weight, exactly as in 2.0.

Spend the Reference Budget Deliberately

Seedance 2.5 accepts up to 50 references per early docs: 30 images, 10 video clips (2 to 30 seconds each, 30 seconds combined), and 10 audio clips (same limits). You address them in the prompt exactly like 2.0: @image1, @video1, @audio1.

The mistake the big budget invites is stuffing. Fifty slots doesn't mean fifty opinions; every reference is an instruction, and contradictory references fight each other. Think of the budget in layers:

• Identity layer (3 to 6 images): the character from multiple angles. A turnaround sheet (front, profile, three-quarter) beats six random selfies.
• World layer (2 to 5 images): environment plates, key props, wardrobe.
• Style layer (1 to 3 images): a color-grade frame or film still that defines the look.
• Motion layer (0 to 2 videos): a clip whose choreography or camera movement you want followed.
• Sound layer (0 to 2 audio): the song to cut to, or a voice reference.

Two rules carry over from 2.0 and matter even more now:

1. Order is priority. The model weights early references more heavily, so @image1 should be the thing you most need preserved (almost always the face).
2. Name the reference's job in the prompt. "The woman from @image1 wearing the jacket from @image3, in the alley from @image5, moving like @video1, cut to the rhythm of @audio1."

Prompt the Audio, Don't Just Enable It

In 2.5, audio is co-generated with the picture in one latent space, which means the prompt's audio language actually shapes the edit. Treat sound as a first-class prompt layer:

• Diegetic detail: "rain on the car roof, wipers squeaking, muffled jazz from the radio" beats "generate audio."
• Dialogue with delivery: write the line and how it's said: "she says, almost laughing: 'you're late.'"
• Music as structure: with an @audio1 song reference, tell the model what syncs: "cuts land on the beat, final pose on the last hit."

Lip-sync is meaningfully tighter in 2.5, so dialogue scenes that were risky on 2.0 are worth attempting.

Camera Vocabulary Still Wins

Precise camera language remains the highest-leverage skill in Seedance prompting, and 2.5 responds to more of it: rack focus, crane moves, and whip pans join the reliable push-ins and steadicam follows. Our camera movements library has copy-paste phrasing and examples for 54 moves, and all of it applies here.

One habit worth keeping from 2.0: avoid the bare word "fast." It remains the keyword most likely to degrade motion quality; describe speed concretely instead ("whip pan," "explosive burst," "snaps into position").

The Parameters Worth Knowing

Beyond the prompt, early 2.5 API docs list a few controls that change results:

• duration — 4 to 30 seconds; ask for what the story needs, not the maximum
• generate_audio — defaults on; disable only if you're scoring in post
• camera_fixed — set true for locked-off shots (product work, dialogue two-shots); it beats writing "static camera" in prose
• seed — fix it to iterate one variable at a time between generations
• return_last_frame — hands you the final frame as a clean image, the raw material for chaining a follow-up clip
• Resolution: iterate at 480p, finalize higher; per-second billing makes cheap drafts the whole game

A Full 2.5-Style Prompt, Assembled

The man from @image1 (identity), wearing the suit from @image2, walks through the hotel lobby from @image4. Style and grade follow @image6, moody amber tungsten. 0-8s: steadicam follow from behind as he crosses the marble floor, staff turning to watch. 8-18s: he pushes through brass doors into the night; whip pan to reveal a waiting crowd of photographers, flashbulbs strobing. 18-30s: slow push in on his face, half-smile, rack focus to the marquee behind him. Cut the edit to @audio1, beats landing on the flashes. Diegetic sound: crowd murmur, camera shutters, distant traffic. Cinematic film grain, 24fps.

Every element is doing declared work: identity pinned to a reference, wardrobe and location pinned to references, three beats with one job each, camera moves named precisely, music synced explicitly, audio described diegetically.

Common Mistakes to Avoid

• Writing a 30-second clip as one unbroken paragraph. Beats or mush.
• Uploading references without assigning jobs. Unlabeled inputs get averaged, and averaging is where identity drift comes from.
• Contradicting your references in prose. If @image2 is a navy suit, don't write "black tuxedo." The model will split the difference and lose both.
• Maxing duration by default. Per-second billing means an unnecessary 30-second generation costs double the 15-second version of the same idea.
• Treating audio as a checkbox. Silent-movie prompts with audio enabled produce generic ambience; described sound produces sound design.

Frequently Asked Questions

Do Seedance 2.0 prompts work on 2.5? Yes, unchanged. 2.5 adds structure on top (beats, roles, bigger reference budgets) rather than new syntax. See our full 2.5 vs 2.0 comparison for what's different under the hood.

How many references should I actually use? As few as pin down what you can't afford to lose. A strong generation usually needs 4 to 8, not 50. The big budget exists for productions with many characters, products, and scenes, not as a target.

Where can I practice these prompts? Seedance 2.0 today, anywhere it runs; the skills transfer directly. Kie.ai runs Seedance 2.0 cheaply now and has confirmed 2.5 API support with free credits for new users at launch.

Skip the Prompt Engineering Entirely

Starrd's templates are pre-engineered Seedance prompts refined over hundreds of generations: beat structure, reference ordering, camera grammar, audio direction, all handled. Upload a photo or two, pick a scene, and star in the result.