Seedance 2.5 Prompt Guide: Examples & Templates

A strong Seedance 2.5 prompt reads less like a list of visual adjectives and more like a compact production brief. It identifies the subject and visible event, assigns a job to every reference file, divides longer scenes into stages, and states what must remain unchanged. That is the through-line of ByteDance's official Dreamina Seedance 2.5 Prompt Guide.^[1]

This article turns that guide into a practical reference library. Every worked prompt is an original practical template or a clearly labeled adaptation of an official teaching scenario; none is presented as generated output.

This is specifically a Seedance 2.5 prompt guide. If you need a tour of modes, upload controls, and generation settings instead, use the separate Seedance 2.5 user guide. Keeping prompt design and product operation separate makes both resources easier to use.

The answer in 60 seconds

• Start with subject + action/event. Add scene, visual style, camera, and audio only when they change the result.
• Treat every uploaded file as a production department: an image defines appearance, a video defines motion or camera, and audio defines voice, music, rhythm, or ambience.
• Bind references to named subjects one by one. Never expect the model to infer that “images 1–4 respectively” means four different people.
• For a long prompt, write consecutive stages with one primary state change and a visible end state per stage. Use precise seconds only for events that truly need a cue.
• When editing, declare the source video the sole timeline and camera master. Describe the edit scope, the target reference's role, and everything that must be preserved.
• Put music in parentheses, sound effects in angle brackets, dialogue in braces, and subtitles in full-width brackets when those distinctions matter.
• End with a consistency block covering identity, count, clothing, prop ownership, spatial direction, audio, and exclusions.

The reusable skeleton is:

[Generation goal]
[Materials]
@Image 1 defines ...
@Video 1 defines ...
@Audio 1 defines ...

[Scene or stage plan]
Subject + visible action/event + environment + style + camera + audio.

[Maintain consistency]
Keep ... unchanged. Do not introduce ...

First, separate Dreamina 2.5 limits from this site's current API

The official guide describes the upstream Dreamina Seedance 2.5 authoring surface. It does not mean every third-party product or API exposes the same limits on the same day. Check the route you are actually on before copying a 30-second, 50-reference prompt into another interface — on this site the Seedance 2.5 route now matches those upstream ceilings, while the older 2.0 routes do not.

The Dreamina limits come from the two official 2.5 documents.^[1][2] For stability, the guide recommends 1–8 distinct image subjects; 1–5 video subjects with roughly 5–10 seconds per subject; and, for editing, a source under 20 seconds with 1–5 supporting references. Higher counts can work but are less stable. The Seedance2.so column reflects the current studio configuration, not the model's theoretical ceiling.^[4] Since the Seedance 2.5 route launched here, its material ceilings and 30-second output match the documented upstream numbers, so the practical gap now shows up mainly on the older 2.0 routes. Seedance2.so is independent and not operated by, affiliated with, or endorsed by ByteDance.

Examples below use the official labels @Image 1, @Video 1, and @Audio 1. In this site's reference-to-video tool, use the compact field labels shown by the UI: @image1, @video1, and @audio1.

The official Seedance 2.5 prompt formula

ByteDance gives a six-part formula:^[1]

Subject + Action/Event + Scene/Environment (optional) + Visual Style (optional)
+ Camera Movement/Cut (optional) + Audio (optional)

Only the subject and event are essential. Include an optional field when it makes a decision. If an image fixes the room, repeating the wallpaper wastes attention; if the camera move defines the commercial, name it.

The official guide teaches its basic formula with a ceramic-studio scene. Here is a more production-ready adaptation of that teaching scenario:

Create an 8-second realistic video set in a quiet pottery studio just after sunrise.

[Subject and goal]
One ceramicist completes one pale-blue cup on a spinning wheel, lifts the same cup with both hands, and places it upright at the center of a wooden display shelf.

[Timeline]
0-3s: Medium eye-level shot. The ceramicist makes the final smoothing pass around the wet rim while the wheel turns slowly.
3-5s: Gently push in to show the damp glaze, small finger marks, and the circular motion stopping. The ceramicist lifts the single cup; keep its shape and color unchanged.
5-8s: Cut to a frontal shelf view. The ceramicist places the cup on the center mark, releases both hands, and holds for one second on the completed arrangement.

[Light, camera, and sound]
Use soft dawn light through one side window, natural wood tones, shallow but readable depth of field, and restrained documentary realism. Preserve the low wheel motor, wet-clay friction, one quiet ceramic-on-wood contact, and subtle room tone. No dialogue or music.

[Continuity]
Exactly one ceramicist and one pale-blue cup. Keep the bench orderly. Do not duplicate the cup, change its color, add text, skip the shelf placement, or end before both hands release it.

The first line supplies subject, action, object, and time. The rest specifies light, one camera move, sound, and the object that cannot drift—without padding the prompt with empty quality adjectives.

Prompt settings are not scene prose

Aspect ratio, output duration, and resolution belong in the generation controls or API request. The official guide explicitly tells creators not to spend prompt space restating settings already selected on the page.^[1] Write vertical phone video only if you want the visual language of a phone recording, not merely because the selected canvas is 9:16.

Some workflows also lock parameters. Editing approximately follows the source video's ratio and duration; a first-frame or first-and-last-frame job inherits the first image's ratio; extension follows the source ratio. In those cases, prose cannot override the locked input geometry.^[1]

Build a reference map before writing the scene

The official multi-reference workflow is simple to state and easy to skip:

Define each material's role → Map subjects → Group by type
→ Create subject profiles → Select references for each scene

This order prevents the model from trying to use every reference everywhere.

1. Give each material one primary job

Use explicit role lines:

@Image 1 defines Mara's face, short black hair, and green work jacket.
@Image 2 defines the brass desk lamp: its proportions, patina, and pull chain.
@Image 3 defines the archive room's shelving, colors, and window placement.
@Video 1 defines Mara's hand movement and the slow left-to-right camera track.
@Audio 1 defines the quiet room tone and the cadence of the spoken line.

Then state exclusions when a file contains information you do not want inherited:

Use @Video 1 for motion and camera only. Do not inherit its actor, clothing,
background, lighting, subtitles, or music.

This is more reliable than “make a video based on all references.” The latter never says which source wins when two files disagree.

2. Bind every subject individually

Avoid compact mappings such as images 1–4 are the four characters respectively. They force the model to infer upload order, identity, and scene participation at once. Bind each one:

@Image 1 defines Nia, the courier in the orange raincoat.
@Image 2 defines Sol, the florist in the navy apron.
@Image 3 defines the paper parcel carried only by Nia.
@Image 4 defines the bicycle parked outside; neither character rides it in this scene.

If several angles show one product or person, say so. The official guide notes that separate views are often more stable than a collage:^[1]

@Image 1, @Image 2, and @Image 3 are three views of the same single espresso machine.
Together they define one machine's shape, materials, controls, and logo placement.
Show exactly one machine. Do not create three machines or a split-screen collage.

3. Group a large pack by production role

For a complex scene, group the reference ledger before the plot:

[Characters]
@Image 1 defines Nia. @Image 2 defines Sol.

[Props]
@Image 3 defines Nia's parcel. @Image 4 defines the flower-shop register.

[Scenes]
@Image 5 defines the storefront exterior. @Image 6 defines the packing counter.

[Motion and audio]
@Video 1 defines the parcel handoff and camera path only.
@Audio 1 defines street rain and the bell above the door only.

4. Create a subject profile, then choose references per scene

A reusable profile keeps identity separate from plot:

[Nia profile]
Appearance: @Image 1. Clothing: orange raincoat from @Image 1.
Fixed prop: the parcel from @Image 3 remains in Nia's right hand until the handoff.
Motion reference: walking pace from @Video 1.
Exclusions: no hat, no umbrella, no wardrobe change, no duplicate parcel.

For each scene, list only the material needed there:

[Scene 1 — storefront]
Use: @Image 1, @Image 3, @Image 5, @Audio 1.
Event: Nia approaches the closed flower-shop door through light rain.
End state: her right hand reaches the brass handle; the parcel remains under her left arm.

[Scene 2 — counter]
Use: @Image 1, @Image 2, @Image 3, @Image 4, @Image 6, @Video 1.
Event: Sol receives the parcel and places it beside the register.
End state: Sol owns the parcel; Nia's hands are empty.

The end state becomes the next scene's starting truth; without it, props and character counts drift.

Here is a complete practical template:

Mode: Omni Reference / Reference to Video
Materials: @Image 1 Nia; @Image 2 Sol; @Image 3 parcel; @Image 4 register; @Image 5 storefront; @Image 6 packing counter; @Video 1 handoff motion; @Audio 1 rain and door bell
Parameters: 15 seconds; 16:9; 720p

Prompt:
[Characters] @Image 1 defines Nia in one orange raincoat. @Image 2 defines Sol in one navy apron.
[Props and scenes] @Image 3 defines the only parcel. @Image 4 defines the register. @Image 5 defines the exterior. @Image 6 defines the counter.
[Motion and audio] @Video 1 defines only handoff timing and a slow camera track; do not inherit its people, clothes, room, captions, or music. @Audio 1 defines rain and the door bell only.
[Scene 1 — approximately 0–7s] Nia approaches the storefront through light rain with the parcel under her left arm. End with her right hand on the brass door handle.
[Scene 2 — approximately 7–15s] Continue inside. Sol receives the same parcel and places it beside the register. End with Sol owning the parcel and Nia's hands empty.
[Consistency] Exactly two people and one parcel. Preserve identity, clothing, shop layout, screen direction, weather, and prop ownership. No bicycle use, subtitles, logo, duplicate parcel, or wardrobe change.

Audio punctuation that gives each sound a role

The official prompt guide uses a compact notation system:^[1]

For non-Chinese speech, name the language, regional variety, delivery, and speaker before the line:

Spanish, Mexican variety, warm but out of breath, spoken by Nia:
{I thought the rain would stop me.}

Do not ask the model to infer the speaker from proximity. Do not leave irrelevant music or subtitles inside a reference file and assume they will be ignored. Either remove them before upload or state that the reference supplies motion only and its BGM, captions, voices, and ambience must not be inherited.

Write 30-second prompts as consecutive state changes

For the official 2.5 workflow, a 30-second scene should not be a paragraph containing twelve actions. Divide it into stages. Each stage gets one primary change and an observable end state.^[1]

Here is an expanded adaptation of the guide's flower-shop packing scenario:

Mode: Dreamina Seedance 2.5 Omni Reference / 30-second generation
Materials: @Image 1 Sol; @Image 2 Nia; @Image 3 flower-shop interior; @Video 1 parcel handoff; optional @Audio 1 room tone
Parameters: 30 seconds; 16:9; 720p

Prompt:
[Generation goal]
A grounded 30-second short film about a florist preparing the final order of the day.
Naturalistic performance, warm shop interior, rain-dark street outside.

[Reference roles]
@Image 1 defines Sol's identity and work clothes. @Image 2 defines Nia's identity and orange
raincoat. @Image 3 defines the shop layout, materials, and lighting. @Video 1 defines only the
parcel-handoff timing and body mechanics; do not inherit its people, clothing, setting, captions,
or music. If supplied, @Audio 1 defines shop room tone only; do not inherit voices or music.

[Stage 1 — approximately 0–8s]
Initial state: Sol stands at the packing counter; loose flowers and an empty kraft box
are arranged in front of her. Event: she selects three pale peonies and wraps the stems.
Camera: locked medium-wide shot with a very slow push.
End state: the wrapped bouquet rests inside the open box.

[Stage 2 — approximately 8–18s]
Continue from the unchanged box and bouquet. Sol folds the box flaps, ties one blue ribbon,
and presses a small blank card beneath the knot. <Paper fold, ribbon tension, rain on window.>
End state: the box is closed and the ribbon is secure; no loose flowers remain in her hands.

[Stage 3 — approximately 18–26s]
The door bell rings. Nia enters in her orange raincoat. Sol turns, carries the same box
with both hands, and passes it across the counter. Use the handoff timing from @Video 1 only.
At 23s, ownership of the box visibly transfers from Sol to Nia.
End state: Nia holds the box; Sol's hands are empty.

[Stage 4 — approximately 26–30s]
Nia gives a relieved smile and exits. Sol watches from the doorway as the bell settles.
(A two-note musical resolution begins only after the door closes.)
Final state: Sol remains alone inside; Nia and the box are outside.

[Maintain consistency]
Keep faces, clothing, shop layout, box dimensions, ribbon color, screen direction,
rain intensity, and prop ownership consistent. Exactly two people. No subtitles or logo.

Use ranges for budgets, points for handoffs, and relative timing for reactions

• Use a range such as approximately 8–18s to allocate scene time.
• Use one exact point such as at 23s for a handoff, entrance, exit, beat, or transition.
• Use relative language such as one beat after the bell when the response matters more than the clock.

Keep ranges consecutive and non-overlapping. Exact-second instructions are not frame-perfect edits, and too many can make actions disappear. “Three gestures in one second” is an impossible performance brief.

On Seedance2.so today, the Seedance 2.5 route runs the full four-stage version above, because it generates up to 30 seconds. You can pick that route in the Seedance studio and paste the staged prompt straight in. On the 2.0 routes, adapt the method to 4–15 seconds: keep two stages, or generate the longer sequence as connected clips. In either case, duration comes from the selected control, not from the prompt — writing "30 seconds" into a route capped at 15 does not unlock it.

Edit a source video without accidentally regenerating it

For editing, declare the source video the sole master for timeline, scene, camera, motion, occlusion, and event order. Then define the target reference, the edit scope, and everything to preserve.^[1]

[Source]
@Video 1 is the sole master video. Preserve its duration, shot order, camera movement,
actor motion, blocking, occlusion, lighting direction, and original audio.

[Target]
@Image 1 defines only the target object's shape, materials, color, and construction.

[Edit]
Replace only the original object described below. Do not redesign the room or actor.

[Preserve]
Keep all unedited pixels conceptually consistent with @Video 1. No new cuts,
characters, props, text, music, or camera movement.

Adapted template: replace one lamp

Mode: Smart Edit / source-video object replacement
Materials: @Video 1 source clip; @Image 1 replacement lamp
Parameters: preserve source aspect ratio and approximately source duration; preserve original audio

Prompt:
@Video 1 is the sole source and timeline master. It shows one table lamp beside an actor.
@Image 1 defines the replacement lamp only: aged brass stem, green glass shade,
short pull chain, and the exact proportions shown in the image.

Replace only the single lamp in @Video 1 with the single lamp from @Image 1.
The replacement must follow the original lamp's position, scale, perspective, occlusion,
reflections, light interaction, and visibility over time. Preserve the actor, table,
room, camera, edit, motion, dialogue, and ambience. Exactly one lamp; no duplicate shade,
no extra furniture, no change to the actor's hands.

“Replace the lamp” is underspecified. The full version identifies the timeline master, appearance source, object count, and relationships that cannot change.

Use the same structure for background replacement, sound cleanup, color changes, or removing an element. Change only the target and preservation list. The video-editing workspace exposes this source-plus-instruction pattern for currently available Seedance 2.0 routes.

Extend forward or backward by designing the boundary

An extension succeeds when the generated side of the seam prepares the exact state seen on the source side. The two boundary frames should connect naturally, though the guide warns they need not be pixel-identical.^[1]

Backward extension template

@Video 1 is the source video. Extend new content before it.
The last frame of the new segment must connect naturally to the first frame of @Video 1.
Match pose, orientation, prop position, background geometry, spatial relationships,
camera height, composition, lighting, and motion direction at the boundary.
Preserve one continuous instance of every subject. Do not reveal early any object
that first enters later in the source video.

Adapted curator example:

Mode: Extend Video backward
Materials: @Video 1 source clip; @Image 1 curator identity and suit; @Image 2 display case
Parameters: add 6 seconds before source; preserve source aspect ratio; preserve source segment

Prompt:
@Image 1 defines the curator's face and charcoal suit. @Image 2 defines the display case.
@Video 1 is the source: it begins with the curator's right hand already touching the case.

Extend 6 seconds before @Video 1. The curator walks from the left aisle toward the same case,
slows, turns her shoulders toward it, and raises her empty right hand. End the extension with
her fingertips about to make contact in the same pose, direction, framing, and light as the
first source frame. The artifact visible inside the source case must not appear outside it,
and no second curator or second case may appear. Use quiet gallery room tone and measured
footsteps in the new segment, then match the source audio naturally at the boundary.
Do not alter the source segment or its original audio.

For forward extension, invert the logic: the first new frame continues the source's last frame. Make its pose and prop relations the extension's initial state. Use video extension when the selected model supports it.

First/last frames, multi-keyframes, and storyboard grids

These three controls solve different problems.

First and last frame: lock the endpoints

In multimodal mode, explicitly assign the anchors instead of vaguely saying both images are endpoints:

Mode: First and Last Frames in multimodal mode
Materials: @Image 1 first frame; @Image 2 last frame; both use the same 9:16 ratio
Parameters: 8 seconds; 9:16; 720p

Prompt:
@Image 1 is the exact first-frame composition: an unopened red umbrella on wet pavement.
@Image 2 is the exact last-frame composition: the same umbrella open above the same child.

[Timeline]
Approximately 0–2s: hold the opening composition long enough to read the closed umbrella as the
child enters from frame-right. Approximately 2–6s: the child reaches down, picks up the handle,
and opens the umbrella in one continuous action. Approximately 6–8s: let the motion settle into
the exact final-frame composition.

Keep the street, rain direction, clothing, umbrella design, and screen direction consistent.
Use light rain, wet footsteps, and one natural umbrella-opening sound; no dialogue or music.
Do not override either anchor with other references. No second child, second umbrella,
teleportation, reverse entry, or unexplained camera cut.

The output ratio follows the first image, so match the first and last image ratios before upload. Extra references may enrich identity or material, but should not redefine endpoint composition.^[1]

Multi-keyframes: lock ordered visible states

Use @Image 1 through @Image 4 as keyframes in this exact order.
Keyframe 1: sealed coffee bag centered on the counter.
Keyframe 2: the same bag opened; beans pour into the grinder.
Keyframe 3: grounds fall into the portafilter; no bag in frame.
Keyframe 4: finished espresso beside the original bag, label facing camera.

Create one continuous preparation sequence. Each keyframe defines the visible end state
of its stage, not a separate product. Preserve bag design, counter, machine, hand appearance,
light direction, and left-to-right action flow.

Independent images are easier to map than a grid. Keyframes control order and major states, not every in-between frame.

Storyboard grid: communicate shot order, not final pixels

The official guide recommends no more than about 15 clean panels, simple line art, minimal labels, and an explicit reading order.^[1]

Mode: Omni Reference / storyboard-grid control
Materials: @Image 1 six-panel storyboard grid; @Image 2 subject identity; @Image 3 flower-shop location
Parameters: 15 seconds; 16:9; 720p

Prompt:
@Image 1 is a six-panel storyboard read left-to-right, top row then bottom row.
Use it only for story order, approximate framing, subject blocking, and camera direction.
Do not inherit line-art style, panel borders, arrows, labels, placeholder faces, or text.
@Image 2 defines the subject's face, hair, and clothing. @Image 3 defines the shop's layout,
window placement, materials, and warm interior light. Do not let either reference reorder shots.

Shot 1, approximately 0–3s: wide exterior, slow approach to the shop.
Shot 2, approximately 3–5s: close-up, hand reaches for the door.
Shot 3, approximately 5–8s: medium interior, bell rings as the subject enters.
Shot 4, approximately 8–11s: over-shoulder view of the package handoff.
Shot 5, approximately 11–13s: close-up on the blue ribbon.
Shot 6, approximately 13–15s: wide closing frame through the rain-streaked window.

Final treatment: naturalistic 35mm drama, warm interior against cool rain,
restrained cuts, continuous room tone, one natural door bell in Shot 3, and no dialogue,
music, captions, logos, or storyboard graphics in the output.

Coarse and fine blockouts

A coarse blockout is a motion and spatial skeleton. Primitive shapes can define paths, blocking, action timing, camera movement, cuts, lighting changes, and sound cues. Map every geometric placeholder to a final subject and explicitly reject the gray geometry.^[1]

Mode: Omni Reference / coarse-blockout transfer
Materials: @Video 1 primitive blockout; @Image 1 cyclist; @Image 2 delivery box; @Image 3 market street
Parameters: 12 seconds; 16:9; 720p

Prompt:
@Video 1 is a coarse blockout. The blue capsule maps to the cyclist from @Image 1.
The orange cube maps to the delivery box from @Image 2. The gray corridor maps to the
market street from @Image 3. Inherit only timing, movement paths, blocking, camera path,
cuts, and screen direction. Render the final subjects and environment photorealistically.
Use restrained bicycle-chain, tire, and market ambience synchronized to the inherited action;
no dialogue or music. Keep exactly one cyclist and one attached delivery box. Do not inherit
primitive shapes, gray materials, guide lines, labels, viewport UI, or empty scenery.

A fine blockout already has complete modeled structures. Use it when you want to preserve geometry, action, space, camera, and cuts while rerendering final materials, characters, lighting, and style.

@Video 1 is the fine 3D blockout and sole structural master.
Preserve architecture, furniture dimensions, actor blocking, action timing, camera lenses,
camera route, and cuts. @Image 1 defines the final actor; @Image 2 defines oak and brushed
steel materials; @Image 3 defines the evening light. Rerender as a finished live-action scene.
Do not show axes, controllers, camera frustums, path lines, default shaders, or viewport UI.

Clean the fine blockout before upload. A visible path spline can become an invented cable; an axis gizmo can become a prop.

One-click video from an image set

“Turn these images into a video” leaves order, motion, editing, and sound unresolved. Use the official sequence:^[1]

Material roles → Image order → Motion amount → Editing style → Visual treatment → Audio

Mode: Omni Reference / one-click image-set video
Materials: @Image 1 desk pack shot; @Image 2 product in hand; @Image 3 outdoor use; @Image 4 final pack shot
Parameters: 12 seconds; 9:16; 720p

Prompt:
@Image 1 defines the opening product-on-desk composition.
@Image 2 defines the same product in a hand.
@Image 3 defines the outdoor use scene.
@Image 4 defines the final pack shot.
Use the images in this order and treat them as one continuous product story.

[Timeline]
Approximately 0–3s: begin on @Image 1 and add one slow 5-degree product turn.
Approximately 3–6s: match cut by product shape and screen position to @Image 2; use only
slight, anatomically plausible hand movement.
Approximately 6–9s: match cut to @Image 3; add shallow parallax and slow background drift while
the product remains the visual anchor.
Approximately 9–12s: match cut to @Image 4, settle all movement, and hold the final pack shot.

Use restrained match cuts based on shape and screen position.
Final look: warm editorial daylight, tactile paper and metal, high contrast but natural skin.
<Soft room tone becomes outdoor ambience; one clean product click at the final frame.>
Preserve product shape, color, label, orientation logic, and scale across all four beats.
No invented text, warped packaging, extra accessories, abrupt zooms, or frantic motion.

A style-reference video may define rhythm, transitions, subtitle treatment, or music while explicitly excluding its actors and locations.

Transitions, emotion, and professional camera language

Describe the visual bridge in a seamless transition

Do not write only transition from @Video 1 to @Video 2. Name the shared shape, direction, color, object, or motion that hides the seam:

Use @Video 1 before @Video 2. Preserve both source clips outside the transition.
At the end of @Video 1, the camera moves into the courier's orange raincoat until orange
fills the frame. Begin @Video 2 on the same orange field, then pull back to reveal a sunset wall.
Match motion direction, exposure, blur, and transition speed. No flash frame, logo, or extra cut.

Give each seam one mechanism; do not stack four effects into the same half-second.

Direct observable performance, not an emotion label

“She becomes sad” names an internal state but supplies no performance. Write the trigger and visible behavior:

Sol hears the door bell and first assumes the customer returned.
She lifts her chin and starts a smile. One beat later she sees the empty doorway;
the smile stops, her shoulders lower, and her gaze settles on the parcel left behind.
Keep the performance restrained: no tears, collapse, shouting, or exaggerated trembling.

For a multi-stage emotion, connect each change to an event: expectation after a sound, doubt after a pause, relief after recognition. That creates a playable sequence instead of asking for three feelings simultaneously.

Use camera terms as unambiguous instructions

Choose one dominant move per shot. If you need an orbit, crane, dolly, handheld shake, and rack focus, divide the idea into shots and state the cut order.

A pre-submission checklist

Before generating, read the prompt as a production crew that cannot see your intent.

• Goal: Is the deliverable one clear video rather than several competing concepts?
• Subjects: Is every person, product, prop, and geometric placeholder named and counted?
• Material roles: Does every image, video, and audio file have one explicit job and exclusions?
• Scene selection: Does each stage list only the references it actually needs?
• State continuity: Does every stage end in a visible state that the next stage continues?
• Timing: Are ranges consecutive? Are exact timestamps reserved for critical cues?
• Camera: Is there one dominant move per shot, with screen direction preserved?
• Audio: Are speaker, language, delivery, music, effects, and unwanted reference audio clear?
• Editing: Is the source video declared the sole master, with a narrow edit scope?
• Endpoints: Do first/last frames share a compatible ratio and have separate descriptions?
• Exclusions: Have you prohibited duplicates, unwanted text, inherited guides, irrelevant BGM, and style leakage?
• Interface limits: Does the prompt fit the model and duration actually selected, rather than an upstream spec you read elsewhere?

The Seedance prompt generator can produce a first draft. Verify every role and remove contradictions before generating.

Diagnose the output by fixing one layer

Change one layer at a time. If identity is wrong, do not simultaneously change lens, lighting, dialogue, and duration.

Using these prompts through an API today

The official documents describe Seedance 2.5 inside Dreamina, including 30-second generation and larger material packs.^[2][3] They are not proof that this site offers a public Seedance 2.5 developer endpoint; it does not make that claim.

For production access available today, developers can send the same role-mapping principles to the Seedance 2.0 API overview, use reAPI's released Seedance 2.0 model, or test reference behavior in reference-to-video. Adapt the job to whichever envelope that route exposes: the 2.0 routes take up to 9 images, 3 videos, 3 audio files, 15 seconds of reference video/audio, and 4–15 seconds of output, while the Seedance 2.5 route in the studio takes up to 30 images, 10 videos, 10 audio files, and 4–30 seconds of output. The prompt craft transfers; the capability envelope does not.

FAQ

What is the best Seedance 2.5 prompt structure?

Start with subject and visible action. Add environment, style, one camera instruction, and audio only where they change the scene. For multimodal work, place a material-role ledger before the scene and a consistency block after it. For longer video, divide the middle into consecutive stages with explicit end states.

How many references can a Seedance 2.5 prompt use?

The official Dreamina 2.5 documentation allows up to 50 total: as many as 30 images, 10 videos, and 10 audio files, with 30 seconds total for video and 30 seconds total for audio.^[1] The Seedance 2.5 route on Seedance2.so exposes the same counts — 30 images, 10 videos, and 10 audio files — while the older 2.0 routes stop at 9 images, 3 videos, and 3 audio files with shorter reference and output limits.^[4] Check the controls on the route you selected rather than assuming one ceiling for the whole site.

Should every uploaded reference appear in every scene?

No. Assign each file a role, build subject profiles, and list the required references per scene. A motion reference can control camera timing without contributing its actor or background. The objective is correct selection, not simultaneous use of every asset.

How should I write timestamps for a 30-second video?

Use consecutive ranges for stage budgets and exact seconds only for crucial handoffs, entrances, exits, transitions, or audio beats. Give each stage one primary state change and a visible end state. Exact timestamps guide important events; they do not guarantee frame-accurate editing.

What is the difference between a keyframe grid and a storyboard grid?

Separate keyframes define ordered visual states more directly. A storyboard grid communicates plot, shot order, approximate composition, and camera ideas. When using a grid, state the reading order and reject its borders, arrows, labels, line art, and placeholder faces.

How do I keep a source-video edit from changing everything?

Declare the source video the sole master for timeline, camera, motion, blocking, occlusion, and event order. Define the target reference's narrow role, state the exact edit and object count, and list everything that must remain unchanged.

Does a prompt unlock 30-second output on Seedance2.so?

No. Duration is a model parameter, not a magic phrase — but the ceiling itself depends on the route. The Seedance 2.5 route here generates up to 30 seconds, matching Dreamina's documented workflow, while the 2.0 routes stop at 4–15 seconds depending on provider and tier. Selecting 2.5 is what buys the longer take; writing "30 seconds" into a prompt on a 15-second route does nothing. Always trust the selected model's visible controls.

Is Seedance 2.5 API access available on this site?

No public Seedance 2.5 developer endpoint is claimed here. Use the current Seedance 2.0 API overview or reAPI's released Seedance 2.0 model as the automation baseline. Verify any future 2.5 API release against an official source before building around it.

Turn the reference map into your next prompt

Start with the smallest job that proves the structure: one subject, one visible state change, one camera decision, and only the references that control those choices. Put every material role above the scene, then put count, identity, prop ownership, direction, and exclusions below it. Every prompt card in this guide is complete and ready to copy; use the surrounding explanation and final checklist to decide what to inspect before accepting a render. The fastest way to test one is to open the Seedance studio, pick the route that matches your material, and run the smallest version of the prompt first.

Use the official 30-second and larger-reference templates in Dreamina when those controls are present. In a smaller third-party or API workflow, preserve the same directing logic and reduce the material pack, duration, and number of stages to the limits shown by that channel. A useful Seedance 2.5 prompt guide should make the next generation easier to diagnose, not merely make the prompt longer.

References

1. ByteDance / Dreamina. Dreamina Seedance 2.5 Prompt Guide. Official formulas, multimodal material limits, role mapping, staged timing, editing, extension, anchors, storyboards, blockouts, transitions, performance, and camera direction. Last modified July 31, 2026. Retrieved August 2026 from bytedance.larkoffice.com/docx/A88jd0B47oAd8zxWp5ycZFMfnxh
2. ByteDance / Dreamina. Dreamina Seedance 2.5 User Guide. Official mode, input, duration, language, long-video, editing, extension, and Clay Renderer documentation. Last modified July 31, 2026. Retrieved August 2026 from bytedance.larkoffice.com/wiki/NjnWwvf4BiFYFLk2RzrcEgaunGf
3. Dreamina. AI creation home. Official product entry linked from the Seedance 2.5 User Guide. Retrieved August 2026 from dreamina.capcut.com/ai-tool/home
4. Seedance2.so. Current omni-reference fields and model configuration. Independent third-party implementation. The Seedance 2.5 route exposes 30 image, 10 video, and 10 audio input fields with 4–30-second outputs; the 2.0 routes expose 9 image, 3 video, and 3 audio fields with 4–15-second outputs. Retrieved August 2026 from seedance2.so/reference-to-video