The Seedance 2.5 Prompting Guide, in English

Super Lewis — Wed, 05 Aug 2026 15:16:24 +0000

Cross-posted from apimodels.app — the canonical version lives there and is kept up to date.

If you have only ever prompted image models, video models feel like a different instrument. You are no longer describing a frame; you are directing time, camera and sound at once, and the failure modes are unfamiliar — subjects that duplicate halfway through, props that change hands, edits that quietly re-roll the whole shot.

ByteDance released Seedance 2.5 on 31 July 2026 and, unusually, shipped a real prompting manual with it, written by the people who trained the model. It is specific in the way vendor docs rarely are: how many reference assets it accepts, how to address each one, how to stage a 30-second clip so it does not collapse, how to change one object in existing footage without disturbing anything else, and a blunt list of nine things it will not do.

This is that manual walked through in English with every template kept. The limits, syntax and stated boundaries are ByteDance's; the worked examples are mine, written to be copy-pasted rather than translated.

Contents

1. Foundations
2. The techniques that matter
3. Advanced usage
4. Checklist before you submit
5. What it will not do
Running these prompts today

1. Foundations

1.1 The base formula

A prompt is assembled from six parts. Only the first two are required; drop any of the rest that you do not need.

FORMULA

subject + action or event
  + scene and environment    (optional)
  + visual style             (optional)
  + camera work or cutting   (optional)
  + sound                    (optional)

What each part is responsible for:

Subject + action or event — who or what is doing what. This is the floor of the whole prompt. Summarise the main process first; add concrete detail only for the key action; never describe the same action twice.

Scene and environment — place, time, weather, spatial relationships, state of the background.

Visual style — light, colour, material, image texture, overall mood.

Camera work or cutting — shot size, camera position, movement, focus target, how shots join.

Sound — dialogue, timbre, ambience, effects, music.

TEMPLATE

<SUBJECT> <main action or event> in <scene and environment>.
The image is <visual style>.
The camera uses <shot size, position, movement or cutting>.
Sound includes <dialogue, ambience, effects or music>.

EXAMPLE

A luthier finishes a violin in a narrow workshop, lifting the body off
the bench and setting it upright in the drying rack.

Late afternoon light rakes through fine sawdust; fresh varnish shows a
deep amber, the bench is orderly, tools laid out in a worn leather roll.

The camera opens on a medium shot of the hands releasing the clamp, then
pushes slowly in on the grain of the top plate, then cuts to the rack
seen straight on.

Keep the scrape of the plane, the click of the clamp, and quiet room tone.

Generation parameters do not belong in the prompt. Resolution, duration and aspect ratio are set on the generation page or through the API.

1.2 With reference assets: prepare them, then assign roles

You can combine up to 50 reference assets in one job. Each type has a hard input range and a narrower recommended range; the recommended range exists to raise stability and is not a statement about capability.

LIMITS

Images    input range   up to 30, each no larger than 4K
          recommended   1-8 subjects

Video     input range   up to 10 clips, 30s total across all clips
          recommended   1-5 subjects, 5-10s per clip

Audio     input range   up to 10 clips, 30s total across all clips
          recommended   keep only dialogue, timbre, ambience or music
                        that the task actually needs

Editing   input range   source video plus reference images together
          recommended   source under 20s, 1-5 reference images

You may go beyond the recommended range: image subjects can extend to 9-12, audio/video subjects to 6-10, editing reference images to 6-8. The guide is blunt about the trade-off — the more assets, the more easily stability drops.

When a subject needs more than five image references and still needs multiple viewpoints, split the viewpoints across separate images. Several independent view images are usually more stable than several views packed into one grid.

Assign a role to every asset. After uploading, state what each asset supplies. When a person, background or composition inside an asset is likely to bleed into the result, also state what not to take from it.

The mapping between assets and subjects must be written into the prompt. Do not rely on text labels burned into the image, and do not leave the model to work out which asset corresponds to which person, prop or scene.

TEMPLATE — asset roles

@image1 supplies <subject>'s <appearance, clothing, structure or material>.
@video1 supplies <action, camera movement or rhythm>.
@audio1 supplies <character or sound type>'s <timbre, dialogue,
        ambience or music>.

<SUBJECT> completes <main action or event> in <scene>.
The image is <visual style>; the camera uses <camera expression>.

EXAMPLE

@image1 supplies <BAKER>'s features, hair and flour-dusted grey apron.
        Do not take the background.
@image2 supplies <BAKERY>'s counter layout, tiled wall and window
        position. Do not take the people in the frame.
@video1 supplies the rhythm of shaping the dough, lifting it and sliding
        the tray in. Do not take the person's identity, clothing or room.

<BAKER> shapes a sourdough loaf in <BAKERY> and slides it into the deck
oven.

The camera records the shaping in a medium shot, then pushes slowly in on
the crust; keep the scrape of the peel, the oven door, and room tone.

When several images are different viewpoints of the same person or product, say so explicitly, and state how many of the object should exist in the result.

EXAMPLE — multiple viewpoints

@image1 defines the front of the same folding lamp.
@image2 defines its left-side structure.
@image3 defines its right-side structure.
@image4 defines its back structure.

The four images together define one single folding lamp; there is only
ever one folding lamp in the finished video.

When a reference video already carries the action, camera work and ordering accurately, the prompt only needs to say what it inherits — do not restate the choreography move by move. Restating it invites a conflict with the asset itself.

A white-box video mainly supplies motion and a spatial skeleton. You still have to write out the subject, scene, action and style you want generated.

1.3 Special characters for sound and text

Natural language works on its own. When you need to separate music, effects, dialogue and on-screen text, four bracket types disambiguate them.

SYNTAX

Music       ( )     (calm piano plays underneath)
Sound fx    < >     <a bell rings in the distance>
Dialogue    { }     {Hello, welcome back}
Subtitle    【 】    【Chapter One: Departure】

To control subtitles or audio, simply write which sound categories to keep and which not to use. Picture content should still be written as positive description — only asset roles, edit scope, and things likely to bleed in need explicit prohibition.

EXAMPLE

No background music; keep only dialogue, ambience and action effects.
No subtitles.
No audio at all.

Reinforcing the dialogue language. When a line is not in Chinese, name the language before it:

The girl says softly in Japanese: {もう大丈夫です}

When the line text is English but the model speaks it in Chinese, or when you need to control a regional variety, be more specific. The recommended formula:

FORMULA

dialogue language + regional variant or accent + delivery
  + speaker + {line}

EXAMPLE

Dialogue language: American English. The girl says it naturally and
colloquially: {I thought you weren't coming.}

Dialogue language: native Los Angeles American English. A young man says
it in natural LA speech: {No way, you actually made it.}

2. The techniques that matter

2.1 Multi-asset work: tell the model which asset each scene uses

Images, video and audio can be combined. Once there are many assets, the job of the prompt is not to cram them all into one sentence — it is to make the correspondence between people, props, scenes, actions and sound unambiguous. Organise it in this order:

ORDER

per-asset roles → subject mapping → grouping by type
  → subject specification → per-scene invocation

Step 01 — name each subject individually. Different people, products and props each get their own binding.

<PERSON A> corresponds to @image1; take appearance, hair and clothing only.
<PERSON B> corresponds to @image2; take appearance, hair and clothing only.
<PROP A>   corresponds to @image3; take structure, material and colour only.
<SCENE A>  references @image4; take spatial layout, architecture and light
           only, not the people in the frame.

Do not write "@image1 through @image4 define four characters". That never says which image is which character.

Step 02 — group the assets by type.

[PEOPLE]
<CONSERVATOR> corresponds to @image1; appearance, hair, clothing only.
<REGISTRAR>   corresponds to @image2; appearance, hair, clothing only.
<INSTALLER>   corresponds to @image3; appearance, hair, clothing only.
<DOCENT>      corresponds to @image4; appearance, hair, clothing only.
The four do not exchange appearance, clothing, actions, positions or lines.

[PROPS]
<SAMPLE BOX>   corresponds to @image5; belongs to <CONSERVATOR> only.
<RECORD BOARD> corresponds to @image6; belongs to <REGISTRAR> only.

[SCENES]
<LAB>     references @image7; space, materials and light only.
<GALLERY> references @image8; space, materials and light only.

[ACTION AND SOUND]
@video1 supplies <CONSERVATOR> opening the sample box; do not take the
        person or the room from it.
@audio1 supplies <DOCENT>'s timbre and the specified line.

Step 03 — write a specification for an important subject. When one person is used across scenes with several assets, collect their definition in one block.

[SUBJECT SPEC: CONSERVATOR]
Appearance and clothing: @image1.
Fixed prop:              <SAMPLE BOX> from @image5.
Appears in:              <LAB> and <GALLERY>.
Action references:       the box-opening in @video1, the sample-placing
                         in @video2.
Do not:                  wear another character's clothing; hold the
                         <RECORD BOARD> or any presentation equipment.

Step 04 — invoke assets per scene. State what the scene uses, what happens, and the state at the end.

SCENE ONE | inspection in the lab
Uses:   <CONSERVATOR>, <SAMPLE BOX>, <LAB>, and the box-opening from @video1.
Event:  <CONSERVATOR> opens the sample box at the bench and inspects the
        sample inside.
Ends:   <CONSERVATOR> stands at the inner side of the bench; the sample box
        stays beside their own right hand, i.e. on the left of frame.

SCENE TWO | registration in the gallery
Uses:   <REGISTRAR>, <RECORD BOARD>, <GALLERY>.
Event:  <REGISTRAR> checks the numbers on the record board beside a case.
Ends:   the record board is still held in both hands by <REGISTRAR>; no
        other character enters the case area.

The goal of multi-asset work is for the model to pick the right asset in the current scene. It is not a requirement that every asset appears on screen at once.

2.2 Thirty-second video: organise events by stage and end state

When there are many events, break the story into consecutive stages. Give each stage exactly one main state change, and write the state that is directly visible on screen when that stage ends.

TEMPLATE — long video

[GOAL]
Generate a <video type>. The core subject is <subject>; the main event is
<story summary>.

[STAGE ONE]
Opens with:   <initial state of people, props and scene>.
Main event:   <one main action or event>.
Ends with:    <person position, prop ownership or picture state>.

[STAGE TWO]
Carried over: <state that must be preserved>.
Main event:   <one main action or event>.
Ends with:    <observable state>.

[STAGE THREE]
Main event:   <closing event>.
Ends with:    <final picture state>.

[KEEP CONSISTENT]
Keep <identity, headcount, clothing, prop ownership, spatial orientation
and sound relationships> stable.

EXAMPLE — florist order-packing walkthrough

[GOAL]
A walkthrough of a florist packing an order. <FLORIST> and <ASSISTANT>
together trim, wrap and hand off a bouquet.

[STAGE ONE]
Opens with:   <FLORIST> behind the bench; loose stems, shears and wrapping
              paper on the surface.
Main event:   <FLORIST> arranges the stems and trims them to length.
Ends with:    the bouquet is held in <FLORIST>'s left hand; the shears are
              back on the right side of the bench.

[STAGE TWO]
Carried over: both keep the same identity and clothing; the bouquet is
              still held by <FLORIST>.
Main event:   <ASSISTANT> unrolls the paper; <FLORIST> lays the bouquet in
              and ties a green ribbon.
Ends with:    the wrapped bouquet lies flat in the centre of the bench, the
              ribbon knot facing the camera.

[STAGE THREE]
Main event:   <ASSISTANT> lifts the bouquet onto the pickup shelf.
Ends with:    the bouquet is alone in the centre of the pickup shelf; both
              stand behind the bench looking at the finished order.

[KEEP CONSISTENT]
Keep <FLORIST> and <ASSISTANT>'s identities, clothing, the orientation of
the bench, the position of the shears and ownership of the bouquet stable.

Timestamps and pacing. For ordinary narrative, prefer stages. Reach for time expressions — in whole seconds — only when a handover, an entrance or exit, a transition or a specific beat needs control. Time ranges allocate story pacing; a time point pins one key event; relative time describes waiting between events.

EXAMPLE

0-5s:   an empty wooden display stand; a hand sets down a white ceramic
        plate; at the end the hand has left and only the white plate is
        on the stand.
5-10s:  the white plate is removed and a clear glass is set down; at the
        end only the clear glass is on the stand.
10-15s: the clear glass is removed and a green ceramic bottle is set down;
        at the end only the green bottle is on the stand.

Time range        0-3s … 3-7s … 7-12s …
Explicit point    At 5s the camera whips left and completes the transition.
Relative time     Three seconds after the button is pressed, the room
                  lights fade out.

Segments must be continuous and must not overlap. A segment is a time budget for an event, not a precise cut point — an action may land slightly before or after a boundary. Too little content in a segment widens the model's room to invent; too much causes over-cutting or dropped story beats. Do not use timestamps to demand a frequency such as three actions in one second.

2.3 Editing, first/last frame and extension: parameters that lock themselves

Video editing, first-frame or first-and-last-frame generation, and video extension automatically lock some generation parameters based on the input.

AUTO LOCK

Video editing
  aspect ratio   automatically keeps the input video's ratio; not settable
  duration       automatically stays essentially the same as the input;
                 not settable. Frame handling can produce a difference of
                 up to about 0.3 seconds.

First / first+last frame
  aspect ratio   automatically uses the first-frame image's ratio. First
                 and last frame should use the same ratio, otherwise the
                 last frame can be stretched.
  duration       settable

Video extension
  aspect ratio   automatically keeps the input video's ratio; not settable
  duration       settable

Locked parameters cannot be overridden on the generation page or through the API. Everything else follows the options currently available.

2.4 Video editing: declare the master, the scope, and what stays

When editing existing footage, first define the source as the sole master, then state the edit target, the scope it applies to, the target asset, and what must be preserved.

TEMPLATE — general edit

[EDIT GOAL]
Edit @video1; across <the whole clip or a specific time range>, perform
<add / remove / replace / adjust> on <picture object, region or sound
category>.

[SOURCE ROLE]
@video1 is the sole editing master, responsible for <people, scene,
action, composition, camera, occlusion relationships, sound and event
order>.

[TARGET ASSET ROLE]
@image1 or @audio1 supplies <the specified attributes of the target object
or sound>.

[EDIT SCOPE]
Process only <object, region, time range or sound category>.

[PRESERVE]
Keep <the picture, action, sound and timing relationships that must not
change> as in @video1.

EXAMPLE

[EDIT GOAL]
Edit @video1; between 4s and 7s only, change the cold blue light on the
right-hand wall to warm amber.

[SOURCE ROLE]
@video1 is the sole editing master, responsible for the people, the room
layout, the action, composition, camera movement, sound and event order.

[EDIT SCOPE]
Adjust only the colour of the light on the right-hand wall and the area it
illuminates; skin tone may change naturally with the ambient light.

[PRESERVE]
Identity, clothing, expression, position, movement, room structure, camera
movement, dialogue and room tone stay as in @video1.

Case A — replacing a subject.

TEMPLATE

[EDIT GOAL]
Edit @video1; change <original object> to <target object> only.

[SOURCE ROLE]
@video1 is the sole editing master, responsible for the original scene,
camera position, camera movement, motion paths, occlusion relationships
and event order.

[TARGET ASSET ROLE]
@image1 supplies <target object>'s <appearance, structure or material>;
do not take <unrelated background, people or composition>.

[EDIT OBJECT AND SCOPE]
Modify only <the explicit object and region>. The count of the target
object across the whole clip is <number>. Do not modify <what must stay>.

[TIMELINE INHERITANCE]
<Target object> inherits every appearance, movement, occlusion and exit of
<original object>: the same timings, durations, paths and changes of speed.
Apart from the object or region explicitly modified above, all other
people, props, scene content, camera movement, cuts and event order in
@video1 stay as they are.

EXAMPLE

[EDIT GOAL]
Edit @video1; replace the yellow folding lamp with the white folding lamp
from @image1 only.

[SOURCE ROLE]
@video1 is the sole editing master, responsible for the desk, the books,
the hand movements, camera position, camera movement, occlusion
relationships and event order.

[TARGET ASSET ROLE]
@image1 is used only for the white folding lamp's appearance, structure
and material; do not take the background, composition or other objects in
the image.

[EDIT OBJECT AND SCOPE]
There is exactly one white folding lamp throughout. Replace only the
original yellow folding lamp; do not modify the books, desk, hands or
background.

[TIMELINE INHERITANCE]
The white folding lamp inherits every appearance, swing of the arm,
occlusion by the hand and exit from frame of the original yellow lamp:
the same timings, paths and changes of speed. Apart from the object
explicitly modified above, all other people, props, scene content, camera
movement, cuts and event order in @video1 stay as they are.

Case B — replacing a background.

TEMPLATE

[EDIT GOAL]
Edit @video1; replace <the original background region> with <the target
environment> from @image1 only.

[SOURCE ROLE]
@video1 is the sole editing master, responsible for the people, foreground
objects, action, composition, camera movement and event order.

[TARGET ASSET ROLE]
@image1 is used only for <target environment>'s spatial layout, materials,
depth of field, ambient colour and light direction; do not take the people
or foreground objects in the image.

[EDIT OBJECT AND SCOPE]
Modify only <the background region outside the subject's outline>. Do not
modify <the subject's identity, features, hair, clothing, expression,
position, size or movement>.

[TIMELINE INHERITANCE]
The subject's movement and occlusion relationships stay as in @video1.
Apart from the object or region explicitly modified above, all other
people, props, scene content, camera movement, cuts and event order in
@video1 stay as they are.

EXAMPLE

@video1 is the sole editing master, responsible for the person, the action,
composition, camera and event order.
@image1 supplies only the spatial layout, depth, ambient colour and light
direction of a daytime glass greenhouse; do not take the people in it.

Replace only the pale grey background outside the person's outline in
@video1 with the daytime glass greenhouse from @image1.

The person's identity, features, hair, clothing, expression, position,
size and raised-hand movement stay as in @video1. Apart from the region
explicitly modified above, all other people, props, scene content, camera
movement, cuts and event order in @video1 stay as they are.

Editing sound. Dialogue, language, timbre, background music and effects can each be handled separately. State the speaker or sound category, the intended change, and whether the other sounds are preserved.

EXAMPLE

Edit @video1; remove the original background music only, keeping the
dialogue, lip sync, ambience and action effects. Picture content, camera
and cutting rhythm stay as in @video1.

Edit @video1; change <DOCENT>'s dialogue language to natural American
English, keeping the line content and the timing of speech unchanged.
Other characters' voices, background music, ambience and picture stay as
in @video1.

2.5 Extension: align the boundary frame first, then describe what comes before or after

Extension generates new material outside the boundary of an existing clip. The output ratio automatically follows the input; the added duration is set by you.

BOUNDARY

          extend backward  ←  [ original video ]  →  extend forward
                              first frame  last frame

The boundary frame must be continuous. Extending forward, the first frame of the new segment has to continue from the source's last frame. Extending backward, the last frame of the new segment has to join the source's first frame. Beyond the boundary frame itself, people, props, background, motion trend and sound must also stay continuous.

Extending forward. Describe the continuous state of the source's last frame first, then what happens after it.

TEMPLATE — forward, basic

@video1 is the source video to extend forward.

Extend @video1 forward. The first frame of the extension continues
directly from the last frame of @video1: keep <subject pose and facing>,
<prop positions>, <background and spatial relationships>, <camera position
and composition>, <light>, <sound state> and <motion trend> continuous.

Then, <describe the new action, event, camera move or sound>.

Throughout the extension keep <identity and clothing>, <key props>,
<background layout>, <the camera axis> and <the original sound
environment> continuous. The same subject remains one continuous object —
no duplication, no splitting; the subject's form and the number of parts
stay stable.

EXAMPLE

@video1 is the source video to extend forward.

Extend @video1 forward. The first frame of the extension continues
directly from the last frame of @video1: keep the same locked-off medium
shot, the position and heading of the orange paper plane, the classroom
window background, the afternoon light and the drift toward the right of
frame continuous.

Then let the orange paper plane glide further right and leave frame, while
the white curtain at the window sways slightly. Camera and classroom
background stay as at the source's last frame.

TEMPLATE — forward, with other reference assets

@image1 defines <PERSON A>'s facial features.
@image2 defines <PERSON A>'s clothing.
@image3 defines <key prop>'s structure and material.
@video1 is the source video to extend forward.

Extend @video1 forward. The first frame of the extension continues
directly from the last frame of @video1: keep <the boundary picture and
sound state> continuous.

Then, <the new action or event PERSON A performs with the key prop>.

Throughout the extension keep <identity and clothing>, <key props>,
<background layout>, <the camera axis> and <the original sound
environment> continuous. The same subject remains one continuous object —
no duplication, no splitting; the subject's form and the number of parts
stay stable.

EXAMPLE

@image1 defines <GARDENER>'s facial features.
@image2 defines <GARDENER>'s pale green work apron.
@image3 defines <WICKER BASKET>'s structure and material.
@video1 is the source video to extend forward.

Extend @video1 forward. The first frame of the extension continues
directly from the last frame of @video1: keep the greenhouse bench,
<GARDENER>'s standing position and the position of <WICKER BASKET>
continuous.

Then <GARDENER> lifts <WICKER BASKET> with both hands and sets it on the
middle shelf of the wooden rack behind them.

Throughout, keep <GARDENER>'s face, apron, the greenhouse layout and the
camera direction continuous.

Other assets may supply people, props or sound, but they do not replace the source's last frame as the control on the opening picture.

Extending backward. Describe what happens before the source starts first, then write the source's first frame as an explicit end state. Writing only "then it joins the original video" tends to make the model introduce later characters or effects early, or keep changing the picture after it has already reached the target state.

TEMPLATE — backward, basic

@video1 is the source video to extend backward.

Extend @video1 backward. Before the source begins, <describe the preceding
action, event, camera move or sound>.

The last frame of the extension joins the first frame of @video1 naturally:
<subject pose and facing>, <prop positions>, <background and spatial
relationships>; keep <camera position and composition>, <light>, <sound
state> and <motion trend> consistent with @video1's first frame.

Throughout the extension keep <identity and clothing>, <key props>,
<background layout>, <the camera axis> and <the original sound
environment> continuous. The same subject remains one continuous object —
no duplication, no splitting; the subject's form and the number of parts
stay stable.

EXAMPLE

@video1 is the source video to extend backward.

Extend @video1 backward. Before the source begins, show the same glass
greenhouse empty: morning mist lying low and slowly dispersing, the shade
blind at the top gradually rising, no people in frame.

The last frame of the extension joins the first frame of @video1 naturally:
keep the central aisle, the planting benches on both sides, the glass
frame, the soft morning light and the locked-off wide composition
consistent with @video1's first frame; at the end the blind is fully
raised, no one is in the aisle, and the leaves are still moving faintly.

TEMPLATE — backward, with other reference assets

@image1 defines <PERSON A>'s facial features.
@image2 defines <PERSON A>'s clothing.
@image3 defines <key prop>'s structure and material.
@video1 is the source video to extend backward.

Extend @video1 backward. Before the source begins, <PERSON A completes the
preceding action or event>.

The last frame of the extension joins the first frame of @video1 naturally:
<PERSON A's pose and facing>, <the position and state of the key prop>,
<where the other characters stand>; keep <background and spatial
relationships>, <camera position and composition>, <light>, <sound state>
and <motion trend> consistent with @video1's first frame.

Throughout the extension keep <identity and clothing>, <key props>,
<background layout>, <the camera axis> and <the original sound
environment> continuous. The same subject remains one continuous object —
no duplication, no splitting; the subject's form and the number of parts
stay stable.
<Assets that appear only after the source begins> must not appear early in
the backward extension.

EXAMPLE

@image1 defines <CURATOR>'s facial features.
@image2 defines <CURATOR>'s dark blue work coat.
@image3 defines <WOODEN DISPLAY CASE>'s structure and material.
@image4 defines the grey work clothes of the two <INSTALLERS>.
@image5 defines the space and light of <PREP ROOM>.
@video1 is the source video to extend backward.

Extend @video1 backward. Before the source begins, <CURATOR> walks to the
bench, picks up the closed <WOODEN DISPLAY CASE> and opens the lid.

The last frame of the extension joins the first frame of @video1 naturally:
<CURATOR> stands in the centre of frame holding the opened <WOODEN DISPLAY
CASE> in both hands; the two <INSTALLERS> stand behind them, one to each
side. Keep the vertical front-on medium shot, the bench position, the prep
room background and the morning light from the left consistent with
@video1's first frame.

The boundary should look naturally joined; it should not be read as pixel-identical. The loudness of the extension may vary slightly from the source. Extending a video generated by the same model generation usually joins more smoothly in both picture and sound. When reviewing, check the picture, the sound and the whole extended segment on both sides of the boundary.

3. Advanced usage

3.1 Keyframes, storyboards and white-box references

First and last frame together with reference assets. In multimodal reference mode you can declare @image1 as the first frame and @image2 as the last frame in the opening line of the prompt — there is no need to switch to a separate first/last-frame mode. The system locks the output ratio to the first frame's ratio; duration is set on the generation page or through the API. First and last frame should use the same ratio, or the last frame may be stretched. Remaining reference images can still define people, props, scenes and materials separately.

TEMPLATE

@image1 is the first frame, defining the composition, subject position,
pose, prop state, scene and camera direction at the start.
@image2 is the last frame, defining the composition, subject position,
pose, prop state, scene and camera direction at the end.
@image3 supplies <SUBJECT A>'s <appearance, clothing, structure or
material>; it does not change the first-frame composition defined by
@image1, nor the last-frame composition defined by @image2.
@image4 supplies <SUBJECT B, prop or scene>'s <specified attributes>; it
does not change the first-frame composition defined by @image1, nor the
last-frame composition defined by @image2.

<Describe one continuous action or event.>
The picture begins naturally from the first frame defined by @image1 and,
through continuous action, arrives at the last frame defined by @image2.
Between the two, keep <identity, prop structure and ownership, scene
layout and camera direction> continuous.

EXAMPLE

@image1 is the first frame, defining the composition of the perfume
workshop at the start: the person's position, pose, the state of the props
on the bench and the camera direction.
@image2 is the last frame, defining the same at the end.
@image3 supplies <PERFUMER>'s face, hair and dark green apron; it does not
change the compositions defined by @image1 and @image2.
@image4 supplies <GLASS BOTTLE>'s shape, material and label position; it
does not change the compositions defined by @image1 and @image2.

Starting from the first-frame pose, <PERFUMER> picks up a pipette and
<GLASS BOTTLE>, drips amber concentrate into it, swirls it gently, seats
the stopper, and places the finished bottle in the centre of the bench,
arriving naturally at the last frame defined by @image2.

Between the two, keep <PERFUMER>'s identity and clothing, the number and
structure of the bottles, the layout of the wooden bench, the warm side
light and the camera direction continuous.

Each anchor image needs its own sentence — do not merge them into "images 1 and 2 as the first and last frames". First and last frame should use the same ratio. Other reference images supply specified attributes only; they do not override the composition of the anchors.

Multiple keyframes in order. When several independent images define different stages of a process, open with the ordering statement, then describe the key state each image represents.

TEMPLATE

Use @image1 through @imageN as keyframes, in that order.

@image1 is the first frame, defining <the composition, subject position,
pose, prop state and camera direction at the start>.
@image2 defines the second keyframe: <the visible state at the end of
stage one>.
@image3 defines the third keyframe: <the visible state at the end of
stage two>.
@imageN is the last frame, defining <the composition, subject position,
pose, prop state and camera direction at the end>.

The picture passes in turn through the states defined by @image1, @image2,
@image3 … @imageN, moving between stages with continuous action.
Throughout, keep <subject identity, prop structure and ownership, scene
layout, light and the camera axis> continuous.

EXAMPLE

Use @image1 through @image4 as keyframes, in that order.

@image1 is the first frame: an orange paper plane at rest on the left of a
classroom desk, nose pointing to the right of frame, locked-off medium shot.
@image2 defines the second keyframe: the same orange paper plane lifted off
the desk by a hand, nose direction unchanged.
@image3 defines the third keyframe: the same plane passing the window, the
curtain swaying slightly to the right.
@image4 is the last frame: the same plane landing on the middle shelf of
the bookcase on the right, nose still pointing right.

The picture passes in turn through the states defined by @image1 to
@image4, keeping flight direction and speed continuous between stages.
Throughout, keep the plane's orange stock, size and fold lines, the
classroom layout, the afternoon side light and the camera axis continuous.

Independent keyframes usually align more easily than several panels packed into one grid image. Understand what they control: stage order and key states — not frame-by-frame reproduction.

Storyboard grids. A grid gives the overall story, the shot order and rough composition. It is not suited to demanding that each cell be reproduced in detail. Keep it under about 15 cells, use clean line art or a tidy schematic, and minimise text labels. State the reading order in the prompt, then write each shot's subject action, shot size or camera move, plus the final look and sound.

TEMPLATE

@image1 supplies the shot order and rough composition of an <N-cell
storyboard>, read <left to right, top to bottom>; do not take <the line-art
style, text labels or placeholder figures> from it.
@image2 defines <SUBJECT A>'s <appearance and clothing>.
@image3 defines <key prop or scene>'s <structure, material or light>.

Shot 1: <shot size, subject action, scene state>.
Shot 2: <shot size, subject action, camera move or cut>.
…
Shot N: <closing action and final picture state>.

The final image uses <visual style>; sound includes <dialogue, ambience,
action effects or music>.

EXAMPLE

@image1 supplies the shot order and rough composition of a four-cell
pottery storyboard, read left to right, top to bottom; do not take the
line-art style or the text labels.
@image2 defines <CERAMICIST>'s face, short hair and dark grey apron.
@image3 defines <BLUE GLAZED CUP>'s proportions, glaze colour and curved
handle.

Shot 1: a wide shot of the quiet pottery studio, <CERAMICIST> seated at
        the wheel.
Shot 2: a medium side shot of <CERAMICIST>'s hands steadying the spinning
        wet clay as the wall rises.
Shot 3: a close-up of fingers refining the join between rim and handle,
        slip running slowly off the fingertips.
Shot 4: a medium close-up of the fired <BLUE GLAZED CUP> being set on the
        wooden rack as <CERAMICIST> withdraws both hands.

The final image is realistic documentary; keep the hum of the wheel, the
friction of wet clay and studio room tone.

White-box references. These split into coarse and fine, and picking the wrong one wastes the asset.

COMPARE

Coarse white-box
  Suits      previsualising action, paths, blocking, camera moves or cuts
             with simple geometry
  Asset      clear relationships between solids, a complete action
             sequence; can be layered with people, prop and scene images
  Prompt     map every white-box solid one by one, and state which
             temporal and spatial information is inherited

Fine white-box
  Suits      a finished model that needs different people, materials,
             colour, scene or style
  Asset      complete structure and a clean plate — no trajectory lines,
             axes or camera frustums
  Prompt     keep structure, action and camera; state which attributes
             are to be re-rendered

A coarse white-box is good for locking motion paths, direction of travel, where people stand, entrances and exits, camera paths, cut points, light changes and sound rhythm. Every solid should map to a final subject or prop; appearance comes from other reference images.

MAPPING — what the white-box supplies → what you must state

Blocking   motion paths, direction of travel, where people stand,
           entrance and exit order
Camera     position, camera path, direction and changes of speed
Light      light direction, changes in brightness, and when they happen
Cuts       cut points, and the subject and composition either side
Sound      whether dialogue, music, ambience or action effects are
           inherited

Prefer simple solids with clear relationships. Appendages such as arms or wings are only suitable as white-box information when the action sequence is complete; otherwise they tend to produce stiff movement or misread structure.

TEMPLATE — coarse white-box

@video1 is a coarse white-box reference. It supplies only <motion paths,
blocking, camera position, camera movement, cuts, light changes, sound
rhythm or spatial relationships>; do not take the white-box appearance,
materials or scene from it.
The <white-box subject A> in @video1 corresponds to <SUBJECT A>.
The <white-box subject B or geometric prop> in @video1 corresponds to
<SUBJECT B or key prop>.
@image1 defines <SUBJECT A>'s <appearance, clothing or structure>.
@image2 defines <SUBJECT B, key prop or scene>'s <specified attributes>.

<SUBJECT> completes <the main action or event> in <scene>.
Keep <the motion paths, blocking, camera movement, cuts, light or sound
rhythm> from @video1.
The final image uses <people, scene, materials and visual style>; sound
includes <dialogue, ambience or action effects>.

EXAMPLE

@video1 is a coarse white-box reference. It supplies only the walking path,
the direction the cart travels, the locked-off camera position, one push-in
and two cuts; do not take the grey solids' appearance or the empty scene.
The tall cylinder in @video1 corresponds to <DOCENT>.
The long box in @video1 corresponds to <MOBILE CART>.
@image1 defines <DOCENT>'s face, blue uniform and name badge.
@image2 defines <MOBILE CART>'s white metal frame and clear cover.
@image3 defines the curved wall, grey floor and linear ceiling lights of a
science gallery.

<DOCENT> pushes <MOBILE CART> along the curved wall, stops in front of the
central plinth and opens the clear cover.
Keep the walking path, blocking, push-in direction and cut points from
@video1.
The image is bright, realistic museum-documentary; keep footsteps, the
sound of the wheels and gallery room tone.

A fine white-box already has complete people, props or scene structure, and suits changing materials, colour, cast, scene and overall style. The plate should be clean — no trajectory lines, axes, controllers or camera frustums.

TEMPLATE — fine white-box

@video1 is a fine white-box reference. Keep <subject structure, action,
spatial layout, camera position, camera movement and cuts>; do not take
the original grey-model materials or the empty background.
@image1 defines <subject>'s <cast, material, colour or surface detail>.
@image2 defines <scene>'s <space, materials, light or visual style>.

Re-render the <subject> in @video1 as <final subject>, and the scene as
<final scene>.
Keep <the structure, action, camera and spatial relationships> from
@video1; the image shows <materials, colour and style>; sound includes
<ambience, effects or music>.

EXAMPLE

@video1 is a fine white-box reference. Keep the complete structure of the
ring assembly, the rotational relationship between the three rings, the
plinth position, the orbiting camera move and the cuts; do not take the
original grey-model materials or the empty background.
@image1 defines the brushed brass material of the outer ring.
@image2 defines the translucent blue glass material of the inner blades.
@image3 defines the white curved wall, dark grey floor and soft ceiling
light of a contemporary art gallery.

Re-render the ring assembly in @video1 as a kinetic sculpture of brass and
blue glass, and the scene as the contemporary art gallery.
Keep the structure, rotation rhythm, orbiting camera move and cuts from
@video1; keep the low hum of the mechanism and quiet interior room tone.

3.2 One-click film

One-click film assembles several images — or images plus a style reference video — into a finished piece with consistent pacing and packaging. The prompt has to state each asset's role, the image order, how much the picture moves, the cutting rhythm, the visual packaging and the sound. Do not just write "turn these into a video".

ORDER

asset roles → image order → amount of motion → cutting style
  → visual packaging → sound

TEMPLATE

[ASSET ROLES]
@image1 supplies <person, product, scene or opening shot>.
@image2 supplies <person, product, scene or process shot>.
@image3 supplies <person, product, scene or closing shot>.
@video1 supplies <cutting rhythm, transitions, caption packaging or music
        style> only; do not take the identities or locations in it.
        (optional)

[ARRANGEMENT]
The images appear in <upload order / a specified order / freely arranged
by theme>.
<State the person, product, place and event relationships to preserve.>

[PICTURE MOTION]
Each image uses <a slight live effect, parallax, push/pull, lateral move
or localised motion>.
Keep <subject appearance, product structure, text or background
relationships> stable.

[FINISHED STYLE]
Use <cutting rhythm, transition style, caption or graphic packaging,
colour style>.

[SOUND]
Include <dialogue, ambience, effects or music>.

EXAMPLE — night-market travel short

[ASSET ROLES]
@image1 supplies the market entrance and the opening environment.
@image2 supplies <TRAVELLER> walking along the street.
@image3 supplies the lantern stall and craft details.
@image4 supplies three friends around a table eating.
@image5 supplies the riverside at night and its reflections.
@image6 supplies the closing shot of the three on the bridge.
@video1 supplies the brisk cutting rhythm, hand-drawn stickers and
        transition style only; do not take the identities or locations.

[ARRANGEMENT]
The images appear in the order @image1 to @image6, forming the sequence
arrive → wander → eat → walk → group photo.
Keep the three friends' appearance and clothing stable; do not blend them
into each other.

[PICTURE MOTION]
Environment images use a slow push-in and slight parallax; images with
people add only natural blinking, head turns, raised cups and clothing
moving in the wind.
Keep the stall structure, table positions and the bridge railing stable.

[FINISHED STYLE]
Use a bright travel-short rhythm; join scenes with natural wipes and
related colours; hand-drawn stickers appear only at the edge of frame.

[SOUND]
Keep market crowd noise, the clink of tableware and river wind, with light
instrumental music that does not dominate.

When image order matters, write the order image by image. When you are happy for the model to arrange them, say so explicitly — "may be arranged freely by theme". With several people or several products, name and bind each one as usual.

3.3 Seamless transitions between two videos

A seamless transition generates continuous material between two clips. State which clip comes before and which after, then the trigger action, camera movement, how the picture changes, the state it arrives at, and how the sound joins.

ORDER

before-clip → after-clip → trigger action → camera movement
  → picture morph → arrival state → sound

FIVE WAYS — what each one needs stated

Dive or whip-back    camera direction, changes of speed, and when the
                     next scene is entered
Subject rotation     the pose, the direction of rotation, and how clothing
                     or background changes continuously
Foreground wipe      when the occluder fills the frame, and the
                     composition revealed after it
Object morph         the corresponding shapes and materials before and
                     after, and the morphing process
Push/pull or focus   camera movement, the focus target, and the
                     continuous spatial relationship between the two

TEMPLATE

@video1 is the before-clip; take its <closing subject, action,
composition, camera direction and sound>.
@video2 is the after-clip; take its <opening subject, composition, camera
direction and sound>.
Keep the <identities, product structure, scenes and main actions> in the
original @video1 and @video2 stable.

At the end of @video1, <subject or foreground object> triggers the
transition through <action>.
The camera <direction and change of speed>, and the <shape, material,
light or space> in frame gradually becomes <the corresponding element> at
the start of @video2.
The transition ends arriving naturally at @video2's opening composition,
keeping <subject position, camera direction and motion trend> continuous.
Sound moves smoothly from <the first clip's sound> to <the second clip's
sound>.

EXAMPLE

@video1 is the before-clip; take the rainy night street, the red umbrella,
the slow push-in and the rain.
@video2 is the after-clip; take the circular skylight of the gallery, the
rising camera and the quiet interior reverb.
Keep the people, the street, the gallery structure and the main actions in
both original clips stable.

At the end of @video1, the red umbrella moves toward the lens and
gradually covers the whole frame, triggering the transition.
The camera continues pushing forward; the circular edge of the umbrella
gradually becomes the metal ring of the gallery skylight, and the red
canopy gradually becomes the white daylight coming through it.
The transition ends arriving naturally at @video2's opening low-angle
composition, the camera turning smoothly from a forward push to a rise.
The rain fades and passes smoothly into footsteps echoing inside the
gallery.

The aim is continuity of picture and sound. You can ask for the main content of both source clips to be preserved, but a generative transition is not a pixel-identical edit splice.

3.4 Emotional direction and observable performance

Words like "tense", "warm" or "oppressive" give the model an overall direction, and leave the performance wide open. To control acting reliably, name what can be directly seen or heard: eyes, brow, mouth corner, breathing, gaze direction, hand movement. You do not need to list every facial detail. For a single emotional turn, two to four of the clearest signals is usually enough; stage it across multiple beats only when the emotion genuinely turns more than once.

TEMPLATE — single turn, the default

Overall the emotion moves from <starting emotion> to <ending emotion>.
After <trigger event>, <subject> first shows <the immediate observable
reaction>.
Then <eyes, brow, mouth corner, breathing, gaze or hand movement>
gradually <changes>.
Finally <subject> expresses <the target emotion> through <a restrained or
explicit outward sign>.

TEMPLATE — multiple stages, keyed to triggers

On hearing or seeing <the first trigger>, <the subject's first observable
reaction>.
After <the second trigger> appears, <the change in the subject's
expression, gaze or breathing>.
Once <the key information> is confirmed, <the emotion the subject is
trying to contain or hide> gradually surfaces through <observable signs>.
Finally, <the subject's closing action, expression or manner of speaking>.

EXAMPLE

Applause for the end of the performance carries from behind the stage. The
young actor's fingers stop on the programme; her gaze turns slowly toward
the curtain while her shoulders stay tight.

Once the curtain call is confirmed, she lets out a small breath, her
shoulders gradually loosen, a restrained smile reaches the corner of her
mouth and her eyes slowly well up — but she never turns to leave.

3.5 Professional camera terms

Basic camera language and popular moves can go straight into the prompt. When a term is uncommon, open to several readings, or when you need precise control of how the picture changes, also state what it acts on, how the picture changes, and the result you want.

BASIC

Shot size    extreme wide, wide, medium, close, extreme close-up
Movement     push, pull, pan, track, follow, orbit, dive, pull-back,
             tilt-up, handheld shake
Position     low angle, overhead, first person

POPULAR — seven moves and what to add

Oner              the subjects, spaces and event order the camera passes
                  through continuously
Dolly zoom        the size the subject holds, and whether the background
                  is drawn in or pushed away
Aerial            the height, direction of travel, and how much
                  environment to reveal
FPV               the first-person flight or run path, speed and turns
Bullet time       which action is frozen or slowed, and the direction of
                  the orbit
Handheld          who is being followed and how much shake — avoid a
                  "handheld feel" with no subject
Speed ramp        where the action accelerates, decelerates or snaps
                  back, and the final resting state

These can be used as-is; but when several subjects are in frame, still state who the move is built around, where it starts and where it ends.

For terms that are genuinely niche, whose meaning varies across the industry, or that the model may not know, keep the term and translate it into observable change.

FORMULA

term + what it acts on + how the picture changes
  + foreground/background relationship + direction or speed

Rack focus: focus moves smoothly from the foreground leaves to the figure
behind. The leaves gradually soften; the figure's face resolves from
blurred to sharp.

For a precise transition, also write the trigger moment, the occluder, the camera direction, how the cut happens, and the composition or motion trend to preserve afterwards.

FIVE EXAMPLES

1 · Shallow-depth portrait
  <PASTRY CHEF>'s eyes and face stay sharp; the glass jars and lights
  behind fall away into soft round bokeh.

2 · Tracking shot
  The camera moves horizontally at the same speed as <SKATER>; the figure
  stays sharp while the street wall smears horizontally from right to left.

3 · Golden hour
  Warm low-angle sun enters from behind <CLIMBER>'s left; the ridge
  ground carries long shadows.

4 · Natural vignette
  The four corners darken gradually while the central <PIANIST> keeps
  normal brightness and skin tone; no black border appears.

5 · Whip-pan transition
  At 5s the camera whips left; when the foreground bookcase fills the
  frame it cuts to the next scene, and after the cut the camera keeps
  moving left at a similar speed.

Aperture, focal length and shutter values can go into the prompt, but the visible result is usually clearer than a number on its own.

4. Checklist before you submit

The guide closes with fourteen questions. Run them before you spend a generation:

 1  Are the subject and the main action or event stated clearly?
 2  Does every reference asset state what to take AND what not to take?
 3  Is each person, product and prop named and bound to its asset?
 4  Are assets selected per scene rather than all forced on screen at once?
 5  Does each stage of a long video have one main change and a written
    end state?
 6  Are headcount, clothing, prop ownership and spatial relationships
    stable?
 7  Does the edit declare a sole master, the scope, the target count and
    what stays?
 8  Are abstract emotions and camera terms cashed out as something
    directly seen or heard?
 9  Do first/last frames and multi-keyframes each state their own role,
    and do first and last frame share one aspect ratio?
10  Does the storyboard state which structure is inherited?
11  For a white-box, have you decided coarse or fine, and written which
    timing, structure, material and style are inherited?
12  Do editing, first/last-frame generation and extension respect the
    auto-locked ratio and duration rules?
13  For an extension, have you checked the boundary frame, the motion
    trend AND sound continuity?
14  Does one-click film state asset roles, image order, amount of motion,
    cutting style and sound? Does a seamless transition state both clips'
    roles, the trigger, the transition process and the arrival state?

5. What it will not do

Nine stated boundaries. These are the requests that end in disappointment:

1  Timestamps allocate event pacing only; they are not frame-level cut
   points.
2  An edit prompt raises the probability that key events align with the
   source, but cannot guarantee frame-exact overlap.
3  Multi-asset work is about selecting and combining the right assets, not
   about showing them all at once.
4  Subtitles, formulas, signage, product specs and frame-precise timing
   that must be exactly right should be handled with pre-composited assets
   plus generation plus post-production.
5  Video editing locks the input's ratio and essential duration; neither is
   settable, and output duration may differ from the input by up to about
   0.3 seconds.
6  First-frame or first/last-frame generation locks the ratio to the first
   frame; duration is settable. Mismatched first/last ratios can stretch
   the last frame.
7  Video extension locks the input's ratio; the added duration is settable.
   The extension's loudness may differ slightly from the source.
8  In one-click film, if image order or character correspondence matters,
   it must be specified in the prompt.
9  A seamless transition aims at visual and audio continuity; it does not
   mean the two source clips stay pixel-identical.

Running these prompts today

Seedance 2.5 is not open through us yet — verified against our own Volcengine Ark account on 4 August 2026, and we do not list a model whose generate button would fail. Seedance 2.0 is live from $0.092/s and shares most of the grammar above: named subjects, per-asset roles, staged events with end states, native synced audio, up to 9 reference images plus reference video and audio. The formula, the naming discipline, the stage-with-end-state structure and the inheritance clause all transfer. When 2.5 opens, the same key and the same endpoint reach it — you change one model string.

Originally published at apimodels.app, where the same guide is also available in Chinese with Chinese-language templates.

Source: ByteDance's official Seedance 2.5 prompting guide — the limits, syntax, auto-locked parameters, checklist and stated boundaries are theirs; the worked examples and the English phrasing are mine. Status figures were measured on 4 August 2026.

If you want to try this grammar today, Seedance 2.0 shares most of it and runs from $0.092/s.

DEV Community: Super Lewis