Seedance 2.5 Prompt Guide

Seedance 2.5 accepts natural-language direction. Begin with the visible event, then add scene, look, camera, and sound only where those details help control the result.

Subject and event

Name who or what is on screen and what visibly happens.

Scene and environment

Set location, time, weather, spatial relationships, and background state.

Visual treatment

Describe lighting, palette, material, texture, or overall mood.

Camera and cuts

Specify shot size, angle, movement, focus, or how one shot becomes another.

Audio

Direct dialogue, voice quality, ambience, effects, or music.

<Subject> performs <main action or event> in <scene and environment>.
The image uses <visual style>.
Frame the event with <shot size, angle, camera movement, or cuts>.
Audio contains <dialogue, ambience, sound effects, or music>.

A ceramic artist finishes a pale blue cup in a pottery studio at dawn, lifts it from the wheel, and sets it at the center of a wooden shelf.
Soft window light reveals a fine sheen on the wet clay; the workbench stays clean and orderly.
Begin on a medium view of the wheel, push slowly toward the cup's surface, then cut to a frontal shelf composition.
Keep the quiet wheel hum, clay friction, and subtle room ambience.

Expand a short idea in four passes

A production prompt is easier to read when it moves from context to story, then closes with rules that apply to the whole clip.

1. 01Assign the referencesIf files are attached, identify each one and state whether it controls a character, place, movement, voice, or another specific attribute.
2. 02Write the creative briefSummarize the subject, location, central event, visual direction, and any distinctive camera behavior in one sentence.
3. 03Describe the progressionUse a storyline or time ranges. For each beat, cover the visible action and add camera, dialogue, effects, or exclusions only when they matter.
4. 04Finish with global rulesRestate continuity requirements that must hold throughout, plus unwanted elements such as extra subtitles, dialogue, or background music.

Leave out any block you do not need. Resolution, duration, aspect ratio, and other selectable generation settings belong in the UI or API rather than inside the prompt.

Seedance 2.5 can combine as many as 50 reference assets. The hard input ceilings and the stability-oriented recommendations are not the same thing: you can go beyond the recommended range, but the result may become harder to control.

Larger sets can still work—for example, 9–12 subjects in image references, 6–10 subjects in audio/video references, or 6–8 images for an editing task—but stability can decline as the set expands. When more than five subjects need several views, provide separate view images instead of merging all angles into one collage.

Do not ask the model to infer which person, prop, place, movement, or sound an upload represents. Bind each asset in the prompt and add exclusions whenever an unwanted background, person, or composition might leak into the result.

@Image 1 defines <subject>'s appearance, clothing, structure, or material>. Do not use <irrelevant content>.
@Video 1 defines <motion, camera path, or pacing>. Do not use <identity, wardrobe, or setting> from the video.
@Audio 1 defines <speaker or sound type>'s voice, dialogue, ambience, or music.

<Subject> completes <main action or event> in <scene>.
Use <visual style> and <camera treatment>.

@Image 1 defines the ceramic artist's face, hairstyle, and dark green apron. Ignore its background.
@Image 2 defines the wooden workbench, window position, and dawn lighting of the pottery studio. Ignore every person in that image.
@Video 1 supplies only the rhythm of shaping clay, lifting the cup, and setting it down. Do not inherit the demonstrator's identity, clothing, or location.

The ceramic artist completes a pale blue cup, lifts it from the wheel, and places it in the middle of the shelf. Start with a medium shot and move slowly toward the clay texture. Retain wheel, clay, and room sounds.

Multiple views of one object

When several images are angles of the same person or product, say so explicitly and lock the expected object count.

@Image 1 defines the front view of one folding desk lamp.
@Image 2 defines the left-side structure of that same lamp.
@Image 3 defines the right-side structure of that same lamp.
@Image 4 defines the rear structure of that same lamp.
All four images describe a single folding desk lamp. Keep exactly one lamp in the video.

If a reference video already expresses the action, camera path, and timing well, describe only what should be inherited. Repeating the motion in different words can introduce conflict. A blockout clip usually supplies spatial structure and movement, so you still need to define the final characters, setting, action, and look.

Natural language is enough for most prompts. Use the syntax below when music, effects, spoken lines, and subtitles need to be separated clearly.

Reinforce the spoken language

Put the language before the line when it is not Chinese. If the system still chooses the wrong language—or a regional voice matters—state language, variety or accent, delivery, speaker, and dialogue in that order.

The girl says softly in Japanese: {もうすぐ着きます}

Dialogue language: American English. The woman speaks in relaxed, conversational American English: {I didn't expect to see you here.}

Dialogue language: natural Los Angeles English. The young man responds in casual local vernacular: {You really made it all the way out here?}

With a large reference set, the goal is not to squeeze every upload into one sentence. Build a small system that connects subjects, props, environments, movement, and audio.

[Subject Profile: Conservator]
Appearance and clothing: @Image 1.
Fixed prop: <Sample Case> from @Image 5.
Locations: <Conservation Lab> and <Gallery>.
Motion: case-opening from @Video 1; sample-placement from @Video 2.
Exclude: other characters' clothing, <Record Board>, and guide equipment.

Scene 1 | Inspection in the Conservation Lab
Use: <Conservator>, <Sample Case>, <Conservation Lab>, and the opening motion from @Video 1.
Event: <Conservator> opens the case at the workbench and inspects the sample.
End state: <Conservator> remains inside the workbench area; the case rests beside the right hand on the left side of frame.

Scene 2 | Registration in the Gallery
Use: <Registrar>, <Record Board>, and <Gallery>.
Event: <Registrar> checks the record number beside the display case.
End state: <Registrar> still holds the board with both hands; nobody else enters the display area.

Multi-reference prompting is a selection problem: help the model use the right material for the current scene rather than forcing every upload to appear at once.

For a clip with several events, divide the story into consecutive stages. Each stage should carry one main state change and end on a directly observable arrangement of people, props, and space.

[Generation Goal]
Create a <video type>. The central subject is <subject>; the overall event is <brief story summary>.

[Reference Assignments — optional]
@Image, @Video, and @Audio references define <identity, scene, motion, voice, or sound>. State what must not be inherited.

[Global Direction]
Environment and texture: <time, place, atmosphere, and physical detail>.
Visual treatment: <style, depth, palette, and lighting>.
Camera language: <framing, perspective, movement, and cutting rhythm>.
Performance focus: <the behavior or emotion that matters most>.
Exclude: <unwanted dialogue, subtitles, music, actions, objects, or distortions>.

[Stage 1]
Initial state: <people, props, and scene at the opening>.
Primary event: <one main change>.
End state: <visible positions, ownership, and scene state>.

[Stage 2]
Continue from Stage 1: <facts that remain unchanged>.
Primary event: <one main change>.
End state: <observable result>.

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

[Maintain Consistency]
Keep <identity, character count, clothing, prop ownership, screen direction, and audio relationships> consistent.

Use timestamps only when timing matters

Stages are the default. Use ranges to allocate pacing, exact time points for one critical beat, and relative timing for delays between events.

Keep ranges consecutive and non-overlapping. They reserve time for an event; they are not frame-accurate edit points. Too little content creates extra freedom, while overloaded ranges can cause rapid cuts or skipped actions.

Scale the same structure for Long Video mode

Dreamina's dedicated Long Video mode supports a selected duration from 30 to 180 seconds. Put the chosen length and aspect ratio near the beginning, map every reference before the story, then use wider consecutive time blocks for the plot. Close with the visual, audio, and continuity rules that must remain active from start to finish.

Editing, first/last-frame generation, and extension inherit some parameters from the source material. Those locked values cannot be overridden separately in the generation UI or API.

Editing pattern

Name one source video as the sole master. Then define the goal, target reference, edit boundary, and everything that must remain untouched.

[Edit Goal]
Edit @Video 1. Within <the full video or a time range>, <add, remove, replace, or adjust> <object, region, or audio category>.

[Source Video Role]
@Video 1 is the sole editing master. It controls characters, scene, actions, composition, camera movement, occlusion, audio, and event order.

[Target Material Role]
@Image 1 or @Audio 1 defines only <the target attributes>.

[Edit Scope]
Change only <object, region, time range, or sound category>.

[Preserve]
Keep <all visuals, movement, audio, and timing relationships that must not change> from @Video 1.

When a box, brush, arrow, or point marks the edit area, mention the annotation but still name the target. For example: “Inside the marked area, replace the boy's blue jeans with tailored black suit trousers during the first 8 seconds.” The mark locates the region; the text defines the object, change, and timing.

The same structure works for subject replacement, background replacement, lighting changes, dialogue language, music removal, and sound-effect edits. For an object swap, state the exact object count and require the replacement to inherit the original object's motion, occlusion, entry, exit, path, and timing.

Forward and backward extension

A forward extension begins from the source's last frame; a backward extension must end on the source's first frame. Describe the shared boundary state before introducing any new event.

@Video 1 is the source to extend forward.
Continue directly from its last frame. Preserve subject pose and orientation, prop positions, background geometry, camera composition, lighting, and motion direction.
Then <describe the new action, camera treatment, and audio>.
Keep identities, clothing, key props, screen direction, and object count continuous. Do not duplicate or split a subject.

Build a recognizable, believable character

Avoid relying on a generic label such as “cinematic woman” or “handsome man.” Establish identity with a small set of stable, observable traits. For realistic people, combine age and visual identity, natural skin detail, 3–4 facial landmarks, gaze, hair, clothing, body shape, posture, and temperament. The same framework also works for stylized or animated characters.

Identity: <age, cultural or ethnic appearance, and overall facial character>.
Skin: <undertone, complexion, pores, freckles, lines, or other natural texture>.
Face: <3–4 distinguishing features across eyes, brows, nose, lips, cheekbones, or jaw>.
Gaze and emotion: <where the character looks, how focused the eyes are, and the feeling conveyed>.
Hair: <color, length, style, texture, condition, and interaction with wind or movement>.
Wardrobe: <cut, color, garment, fabric, condition, and defining details>.
Presence: <build, posture, framing, action, and overall temperament>.

Physical cues are more useful than abstract emotion alone. Instead of only writing “heartbroken,” direct the lowered gaze, restrained breath, tightening jaw, trembling lips, delayed response, or other behavior that makes the feeling visible on screen.

Describe how a transition actually happens

A transition name is only a starting point. Define the method, continuity safeguards, the visual bridge out of Shot A, and the composition that should be visible when Shot B arrives.

@Image 1 is the first frame. It controls the opening composition, subject position, pose, prop state, scene, and camera direction.
@Image 2 is the last frame. It controls the ending composition, subject position, pose, prop state, scene, and camera direction.
@Image 3 defines <Subject A>'s appearance or clothing without changing either anchor composition.

Describe one continuous event that begins naturally at @Image 1 and resolves at @Image 2. Maintain identity, prop structure and ownership, scene layout, and camera direction between the anchors.

• The subject and primary event are explicit.
• Every reference says what to use and what to ignore.
• Each character, product, and prop has a name and file mapping.
• References are selected per scene rather than forced into every scene.
• Each long-video stage has one main change and a visible end state.
• Identity, clothing, object ownership, count, and spatial direction stay consistent.
• An editing prompt names one master, a narrow edit scope, target quantity, and preserve list.
• Emotion and camera terminology are paired with visible or audible cues.
• First/last frames and keyframes each have one role; anchor images share an aspect ratio.
• Storyboard or blockout prompts state the exact structure and detail level to inherit.
• Locked aspect-ratio and duration rules match the chosen task.
• Extensions check boundary image, motion trend, object continuity, and audio.

Important limitations

• Timestamps budget time; they do not define frame-accurate edit points.
• Editing prompts can improve event alignment but cannot promise frame-by-frame overlap.
• Multi-reference creation selects and combines useful assets; it does not guarantee that every upload appears.
• Exact subtitles, formulas, signs, product specifications, and frame-level timing may still require prepared assets and post-production.
• First/last-frame generation may stretch the last image when the two anchor ratios differ.
• Seamless transitions target perceptual continuity, not pixel-identical preservation of both source clips.

Results vary with source quality, task complexity, selected settings, and the relationships among reference materials. This guide explains prompt-writing technique rather than guaranteeing a particular output.