Seedance 2.5 Prompting Guide (Part 1): Prompt Structure and Reference Generation

Seedance 2.5 does more than understand a scene description. It can assign separate, explicit roles to text, images, video, and audio. Before writing a prompt, decide whether the input locks the output, then specify what must be inherited exactly and what should serve only as a reference.

This guide is divided into two parts:

• This part covers prompt structure, asset mapping, timelines, camera language, blockouts, storyboards, and keyframes.
• Part 2: Video Editing, Extension, and Final Assembly covers instruction-based editing, reference-image editing, audio editing, extension, automatic assembly, and seamless transitions.

Models and Tasks

Seedance 2.5 can generate clips up to 30 seconds long, accept up to 50 image, video, and audio assets in one request, and generate dialogue in more than ten languages. Its main capabilities fall into six groups:

“Locked” and “Reference” Tasks Follow Different Rules

Locked tasks include editing, first/last-frame generation, and extension. The input occupies the output timeline or determines its aspect ratio. The model must adapt to that asset, so some generation settings cannot be changed freely.

Unlocked tasks include ordinary references, storyboards, and keyframes. Assets provide semantic or visual constraints while output duration and aspect ratio remain independent. Separate keyframe images usually align more strictly than a storyboard grid.

Input Recommendations

A Reliable Four-Part Prompt

Asset mapping + one-sentence brief + timeline or shot order + global constraints

1. Asset mapping: In upload order, state whether each image, video, and audio file controls the subject, movement, setting, style, or sound.
2. One-sentence brief: Establish the overall direction with subject + location + event + genre/style + special camera treatment.
3. Timeline or shot order: Describe visuals, actions, camera movement, dialogue, and sound in sequence, keeping content density appropriate for the duration.
4. Global constraints: Add rules that apply throughout, such as camera position, lighting, color, ambience, aspect ratio, no subtitles, or no BGM.

A basic prompt does not need a pile of adjectives. The important thing is to make the action, camera, and sound form a complete eight-second arc.

Realistic nature-documentary look on a forest slope in warm afternoon light. A small, round panda cub tumbles clumsily downhill.

Subject and setting: the cub has fluffy fur and a compact, rounded body. The slope contains moss, clover, soil, small stones, and scattered yellow flowers; tree trunks in the background remain softly out of focus.

0–3 seconds: low-angle medium-wide shot. The panda lies across the incline and begins a slow sideways roll, bending the grass beneath its body while dappled sunlight shifts gently in the breeze.
3–8 seconds: the camera follows slightly down and to the right. The panda gradually stops, settles on its belly, turns its round face toward the camera, presses its front paws into the grass, lifts its head once, then relaxes.

Keep the panda sharp and the background naturally defocused, with subtle handheld breathing. Use only wind, grass rustle, and soft rolling sounds. No music and no subtitles.

Mapping Multiple Assets

Do not rely on names written inside reference images, and do not make the model guess which asset belongs to whom. When the asset package grows, bind every input in a list before describing the story.

Asset mapping:
Images 1–2: Character A's appearance; Audio 1: Character A's voice.
Image 3: Character B's appearance; Audio 2: Character B's voice.
Video 1: Reference only the two characters' action order.
Video 2: Reference only the orbiting camera and editing rhythm.
Image 4: Reference only the warm golden backlight and film color; do not reference the people.

Every asset should answer two questions: Who or what does it belong to? What specifically should be referenced? If only part of an asset should be used, also state what must not be inherited.

Timeline, Camera, and Action

Seedance 2.5 can follow whole-second timestamps. Time ranges should be continuous—0–3 seconds, 3–7 seconds, 7–15 seconds—without unexplained gaps.

• Exact time: “At second 5, whip-pan left and complete the transition with a wipe followed by a natural dissolve.”
• Relative time: “After the lead presses the shutter, freeze the frame for one second, then enter the next shot.”
• Content density: If a time range contains too little, the model will improvise; if it contains too much, actions may be omitted or the edit may become frantic.

Common terms such as extreme wide shot, medium shot, close-up, push-in, pull-out, pan, tilt, low angle, overhead, one-take, aerial, FPV, and dolly zoom can be used directly. For a less common term, describe what the audience should actually see:

Starting at second 4, rack focus: the sharp glass in the foreground gradually softens while the woman in the background moves from blurred to sharp. The focus transition is smooth and the camera position does not change.

Describe the main action first and reserve detailed wording for a few memorable beats. For emotion, use visible facial changes—“eyes redden, the corners of the mouth lower, breathing quickens”—instead of an abstract mood alone.

Blockout Reference: Separate Motion Design from Rendering

A blockout can define shot order, framing, character paths, action, and lighting rhythm. The prompt must state what to inherit and fully describe any subject or setting that lacks a visual reference.

In this example, the blockout controls camera staging while nine images define the character design and visual state at different stages.

Prompt

Use Video 1 only for the complete shot order, camera positions, framing changes, subject paths, and timing. Do not copy the blockout materials or character appearance. Images 2–10 constrain the character design, environment, and color at each corresponding stage.

Generate a 30-second, 16:9 cinematic 3D fairytale animation: a girl launches a toy airplane in her bedroom, enters a fantasy sky, dives toward the ocean and underwater, crosses a space-time rift into space, then returns to the carpet asleep as her father enters the frame.

0–3 seconds corresponds to Image 2; 3–5 to Image 3; 5–8 to Image 4; 8–10 to Image 5; 10–19 to Images 6–7; 19–23 to Image 8; 23–28 to Image 9; 28–30 to Image 10. Preserve Video 1's shot structure exactly, add no new shots, and keep character appearance consistent with each corresponding keyframe.

Reference inputs · 10

Generated result · 2

Prompt

Fully re-render Video 1 while preserving its action paths, camera movement, timing, and framing exactly.

Replace the environment with a cyberpunk night city in deep blue and violet. Dense skyscrapers carry holographic ads and neon reflections while distant aircraft move slowly between them. The subject is a small raccoon in a black stealth suit, carefully crossing a rooftop, its silhouette rim-lit by the city behind it.

No background music. Keep only wind, distant mechanical ambience, footsteps, and fabric movement. Do not generate motion paths, axes, or camera guides.

Storyboards: Control Narrative, Not Pixel-Level Alignment

A storyboard grid is useful for defining shot count, framing, and narrative order, but it leaves room for interpretation. Complete the prompt with action, camera, dialogue, and style. Use separate keyframe images when the composition must align more strictly.

Prompt

Asset mapping: Image 1 defines the order, framing, and camera rhythm of nine shots. Image 2 defines the composition and warm/cool palette of a grassland launch site at dusk. Image 3 defines an elderly woman in a golden-yellow dress. Image 4 defines a tall, weathered teal guardian robot.

Generate a 30-second, 16:9 live-action disaster-film sequence. Use 35mm color-film grain, subtle handheld breathing, and shallow depth of field. Contrast warm golden sunset, cool dusk blue, and explosive orange. No black-and-white, illustration, storyboard line art, toy-like rendering, or plastic CG.

Follow Image 1's nine-shot order: the pair watch the rocket from afar; the robot supports the woman; a close-up farewell; rocket liftoff; explosion in the sky; a tear falls; the woman collapses; the robot bends down and wraps around her protectively; end on a wide shot of the embrace. Dialogue is natural English. Place ambience, wind, and explosion sounds accurately.

Reference inputs · 4

Generated result · 1

Keyframes: Define Image Order More Strictly

Upload every key image separately and in sequence. Begin the prompt by stating, “Use Images 1 through N in order as keyframes.”

Prompt

Use Images 1 through 6 in order as keyframes for a 15-second vertical pixel-art science-fiction short. Keep a deep-blue space background, consistent 16-bit pixel texture, and retro electronic music throughout. Connect all frames through one continuous horizontal camera move.

Begin with the orbital station exterior in Image 1. After the camera passes through a window, the space courier in Image 2 turns toward the lens. In Image 3, she removes a glowing energy core from a storage bay. When the alarm lights activate, she moves through the zero-gravity corridor in the pose shown in Image 4, then leaps across a broken catwalk using the action in Image 5. The camera keeps following and enters the reactor chamber in Image 6, where she inserts the core and the room lights change from red to blue.

Keep her face, orange-and-white spacesuit, energy-core design, and station architecture consistent. Motion must remain continuous. No text, subtitles, flicker, or element deformation.

Reference inputs · 6

Generated result · 1

Preflight Checklist

• Is this a locked or an unlocked task?
• Does every asset have a number, a mapped subject, and one clear responsibility?
• Are the time ranges continuous, with content density appropriate for the duration?
• Are uncommon camera terms followed by a description of the visible result?
• Are action, sound, and transition triggers timed explicitly?
• Are constraints for subtitles, BGM, text, and watermarks clear?

Next: Seedance 2.5 Prompting Guide (Part 2): Video Editing, Extension, and Final Assembly.