Seedance 2 Prompt Guide: Shots, References, and Targeted Repairs
Turn a scene idea into a usable Seedance prompt: give each reference a job, write shots for your version, place the sound, and fix one problem at a time.
On this page

A useful Seedance prompt tells the model what must happen on screen, what each reference controls, and what should remain consistent. Start with the visible story beat, give it a camera position and enough time, then add sound. For a sequence, use Shot 1, Shot 2, and so on. With Seedance 2.5, you can also specify integer-second timestamps; with 2.0, use shot labels rather than relying on precise timecodes. That distinction is documented in the current BytePlus prompt guide. Keep the whole prompt compact: ByteDance publishes no hard character limit, but its Volcengine Ark API reference recommends staying within 500 characters for a Chinese prompt and 1,000 words for an English one.
The workbook below provides adaptable prompts for a product reveal, a small narrative, dialogue, and repairs. The examples are original writing exercises based on documented controls, not clips we generated or performance tests. Copy the example that fits your job, replace its subjects and references, and judge the resulting video against the action you actually requested.
The Seedance 2 Prompt Card
Before writing, finish this sentence: “The viewer should see [an event] happen, and understand [the outcome] by the end.” It gives the clip a destination. “A beautiful cinematic forest” describes a setting; “a mechanical squirrel finds a glowing acorn and carries it home” gives you action and an ending.
A prompt card is a planning aid, not a required input schema:
| Decision | What to write down | Example |
|---|---|---|
| The beat | A visible event and its ending | The squirrel finds the acorn and returns to its tree hollow |
| The subject | A short name plus recognizable traits | Squirrel: copper body, green glass eyes, striped tail |
| Reference jobs | Which attached file supplies which detail | Image1 supplies the squirrel; Image2 supplies the forest |
| The shot | Framing, action, and camera movement | Ground-level tracking as it carries the acorn |
| Sound | Who speaks, when, and what else is audible | Rustling leaves and a soft mechanical click; no dialogue |
| Continuity | Details that should carry between shots | Same acorn, tail pattern, and dusk lighting |

Turn those notes into a readable prompt. The model does not need your private repair prediction as part of the initial scene; keep it beside the prompt so you can compare the output later. Put important reference assignments before the shot list, where they are easy to find.
For a simple clip with no attached assets, you can start here:
A small copper mechanical squirrel with green glass eyes and a striped tail
stands beside a glowing amber acorn in a mossy forest at dusk.
In one continuous ground-level medium shot, the squirrel notices the acorn,
picks it up with both front paws, and carries it into a hollow tree on the right.
The camera tracks alongside the squirrel and settles when it reaches the hollow.
Keep the same squirrel, acorn, and dusk lighting throughout.
Sound: leaves rustle under its feet, with a soft mechanical click as it lifts the acorn.
No spoken dialogue, subtitles, or background music.This prompt has a recognizable subject, an action chain, a camera path, and an ending. It does not need a list of lens brands or twenty mood adjectives to communicate the job.
Choose the Version and Task Before Adding Detail
Use the version your actual app or API exposes. A button labeled “Seedance 2” can be a product label; it does not establish which native model or controls sit behind it.
For the documented BytePlus ModelArk route, the differences that affect prompt planning are:
| Planning question | Seedance 2.0 family | Seedance 2.5 |
|---|---|---|
| How long can one generated clip be? | Up to 15 seconds | Up to 30 seconds |
| How should I describe a sequence? | Shot N; precise timestamps are not supported as in 2.5 | Shot N or integer-second timestamps |
| How many references can the request contain? | Up to 15: 9 images, 3 videos, 3 audio files | Up to 50: 30 images, 10 videos, 10 audio files |
| Can an audio reference be the only reference type? | Requires an image or video reference as well | Audio-only referencing is supported |
| What output resolution is documented? | Regular 2.0 supports up to 4K; Fast and Mini support 480p/720p | 480p, 720p, or 1080p |
These are native ModelArk capabilities, not promises that every consumer interface exposes them. The 2.5 tutorial and model table currently identify the native 2.5 model as dreamina-seedance-2-5-260628. A wrapper may use its own model name. The reference maximum is a ceiling, not a target for your creative plan.
Next choose the job. A fresh scene uses references to guide a new result. An edit changes an existing clip. An extension continues it. A first-frame task begins from a specified image. Those jobs need different inputs even if their prompts mention the same subject.
For example, “Use Video1 for the camera movement” describes a reference. “Replace the blue lantern in Video1 with a red lantern while keeping the rest” describes an edit. “Continue Video1 after the door opens” describes an extension. Write the intended job plainly rather than leaving the model to infer it from a collection of style words.
Seedance Prompt Length: No Hard Character Limit, 1,000 Words Advised
As of November 2, 2026, ByteDance does not publish a hard character limit for Seedance prompts. What it does publish is a recommendation in the content.text parameter of the Volcengine Ark video generation API reference, the China-region counterpart of BytePlus ModelArk: keep Chinese prompts within 500 characters and English prompts within 1,000 words. The reason it gives is attention rather than a rejected request. A long prompt spreads the information out, so the model may skip details, keep only the main points, and return a video with some elements missing. The English BytePlus ModelArk Seedance 2.5 prompt guide gives no length figure at all.
The hard caps you see quoted online come from the apps that run Seedance, and they differ widely. That is why one guide says 3,000 characters and another says 10,000. Here are the figures the services publish, read on November 8, 2026:
| Where you run Seedance | Prompt cap | Source |
|---|---|---|
| Runway, Seedance 2.0, 2.0 Fast, and 2.0 Mini | 10,000 characters | Runway Help Center, "Creating with Seedance 2.0", updated October 30, 2026 |
| Runway, Seedance 2.5 | 15,000 characters | Runway Help Center, "Creating with Seedance 2.5", updated November 1, 2026 |
| invideo AI, Seedance 2.0 | 3,000 characters | invideo's Seedance 2.0 prompt guide, updated October 5, 2026 |
| fal.ai API, Seedance 2.0, 2.0 Fast, and 2.5 text-to-video | No cap in the published schema | fal's OpenAPI schemas for those endpoints, which describe prompt without a maxLength |
| BytePlus ModelArk and Volcengine Ark | No hard cap. Volcengine recommends 500 Chinese characters or 1,000 English words | API references above |
Runway counts characters, not words, so a 10,000-character prompt is roughly 1,500–2,000 English words. That is already past Volcengine's 1,000-word advice. Claims of a 4,000- or 5,000-character "Seedance limit" in forum threads usually describe a single app at a single date. If your box stops accepting text, check that app's own help page rather than assuming the model refuses it.
Most prompts never get close. The longest examples on this page, a three-shot scene with two references and a sound line, run about 140–160 words. If yours is heading toward the ceiling, the extra text usually repeats what a reference already shows. Describe the subject once and let Image1 carry its look, cut mood adjectives that don't change the shot, and move a story beat that doesn't fit into a separate generation or an extension. If an element you asked for disappears from a long prompt's result, shorten the prompt before you rewrite the scene.
The same parameter lists which prompt languages each model accepts. Every Seedance model takes Chinese and English. Seedance 2.5 adds nine more: Spanish, Indonesian, Portuguese, Japanese, Malay, Thai, Arabic, Vietnamese, and Korean. The Seedance 2.0 family adds only Spanish, Indonesian, Portuguese, and Japanese. So if you write in Korean, Thai, Arabic, Vietnamese, or Malay and your interface runs a 2.0 model, write the prompt in English or Chinese instead.
Plan References Before You Plan Shots
Assign each attached asset a clear job and identify the relevant part of it. A reference may supply appearance, location, movement, camera behavior, or voice; it does not have to supply all of those at once.
The native prompt guide binds references by upload order within each asset type. Image1 is the first image, Video1 the first video, and Audio1 the first audio. Adding a second image does not rename the first audio file. Your app may display another token format, so use its actual attachment labels rather than assuming every alias is interchangeable. BytePlus explains reference binding and roles here.
Here is a reference map for the squirrel story. The filenames are illustrative; actually attach the corresponding assets before using the prompt.
| Attachment | Prompt reference | Its job | What it should not contribute |
|---|---|---|---|
copper-squirrel.png, first image | Image1 | Squirrel's body, eyes, and striped tail | Background from the character sheet |
mossy-hollow.png, second image | Image2 | Forest floor, hollow tree, dusk colors | Any extra animals in the image |
side-track.mov, first video | Video1 | Low camera height and lateral tracking | Its subject or soundtrack |
Write the assignment explicitly:
Use the copper squirrel in Image1 as the only animal.
Preserve its green glass eyes and striped tail in every shot.
Use Image2 for the mossy forest and the hollow tree, without copying extra animals.
Use Video1 only for the low lateral camera movement; do not copy its subject or audio.If you only need the two images, remove the Video1 sentence and attachment. Unused references make the assignment harder to inspect. If two images show different tail designs, choose which one controls the tail rather than hoping the model merges them correctly.
In 2.5, the combined duration of input videos must stay within 30 seconds, and the combined duration of input audio must separately stay within 30 seconds. Ten slots do not mean ten 30-second clips fit. Two video references of 18 and 15 seconds total 33 seconds; trim or remove one before submitting. These are documented input limits, not reasons to fill every slot.
Use frame roles when the boundary image matters
“Start like Image1” and assigning an image the first_frame role are different instructions. Semantic reference_image input guides the scene; the first-frame role constrains the start and requires ratio: adaptive on the documented 2.5 route. A first/last-frame pair constrains both ends. Match the pair's proportions, since a different last-frame ratio may stretch its contents. The guide distinguishes these modes.
For a start/end product sequence, attach the images with the appropriate frame roles, then write the motion between them:
Animate a smooth reveal between the supplied start and end frames.
The folded paper lamp opens gradually as the camera moves closer.
Its base stays planted in the same position, and the warm light rises gently.
End on the supplied open-lamp composition. No hands, dialogue, or added text.A single storyboard grid gives a high-level story reference. Separate ordered keyframes can provide closer alignment to individual moments, but neither a grid nor a written timecode guarantees frame-exact cuts.
Check character references before writing around them
For ModelArk 2.0 and 2.5, direct uploads of real human faces in image or video references are unsupported. The documented alternatives include approved digital-character assets, verified and authorized portrait-library assets, and qualifying original outputs from the same ModelArk account and platform within 30 days. Editing, recompressing, or moving a trusted output between accounts can break that eligibility. A consent sentence in the prompt does not replace the asset process. See the portrait-asset guide.
The examples here use objects and stylized mascots so you can work on shots and sound without turning prompt writing into a portrait-enrollment tutorial.
Split the Video Into Shots
Give each shot a beginning, a visible change, and a handoff to the next shot. Those relationships do more for continuity than repeating “cinematic” at the start of every line.
For the squirrel story, a useful 12-second plan is:
| Interval | Visible beat | Continuity to preserve |
|---|---|---|
| 0–3 seconds | Squirrel notices the acorn | Same acorn beside the same root |
| 3–7 seconds | Squirrel lifts it and crosses the moss | Acorn stays in its paws |
| 7–12 seconds | Squirrel reaches the hollow and sets it down | Same tree and dusk light |
For this version, attach only Image1 and Image2 from the earlier reference map; omit the optional Video1. Request a 12-second 2.5 clip and paste this prompt:
Use Image1 for the copper squirrel with green glass eyes and a striped tail.
Use Image2 for the mossy forest, hollow tree, and dusk lighting.
The squirrel is the only animal; the acorn glows amber throughout.
0–3 seconds: Medium shot at ground level. The squirrel stands beside an amber acorn
near a tree root, notices its glow, and bends toward it. Fixed camera.
3–7 seconds: Cut to a side view. The squirrel lifts the acorn with both front paws
and carries it across the moss toward the hollow tree. Track alongside it at ground level.
7–12 seconds: Cut to the hollow's entrance. The squirrel steps inside, gently sets
the acorn down, and pauses beside its warm glow. The camera slowly pushes toward the hollow.
Sound: soft leaves underfoot, a small mechanical click at the lift, quiet forest ambience.
No dialogue, subtitles, or background music.
Keep the squirrel's design and the acorn's shape consistent across all shots.The intervals cover the whole requested duration without a gap. Their boundaries are direction for 2.5, not an editing guarantee. Review whether the carry connects to the arrival, whether the acorn remains visible, and whether the final pause has time to register.
For 2.0, replace the timestamp prefixes with Shot 1, Shot 2, and Shot 3, keeping the sequence within its documented duration. Your timing table remains useful as an editorial plan, but do not present it as precise native timestamp control.
Use Camera and Style Cues Without Overloading the Shot
Choose the camera behavior that helps the viewer understand the action. A fixed shot makes a delicate mechanism easy to inspect. Tracking follows a subject moving through a place. A push-in draws attention to a reveal or decision.
Common terms such as “wide shot,” “close-up,” “pan,” and “slow push-in” provide a clear starting point. If you need an unusual technique, describe what the camera actually does. For a dolly zoom, for example, explain that the camera moves backward while zooming in to keep the subject a similar size as the background changes scale. This follows the official guidance on camera descriptions.
One primary move per shot is a useful editing choice, not a model limit. Combine moves when they serve a coherent action; avoid asking a camera to orbit, zoom, pan, and go handheld without saying when each happens.
A 2.0 product reveal you can adapt
This example needs one product reference image attached as Image1. Request a 10-second clip in your 2.0 interface; the shot labels describe order rather than guaranteed cut times.
Use Image1 for the folded paper lamp's cream shade, brass base, and overall proportions.
The lamp stands alone on a pale wooden desk against a soft gray background.
Shot 1: Medium shot of the folded lamp, unlit. Fixed camera. A narrow band of morning
light falls across the desk, leaving the lamp's outline easy to see.
Shot 2: Cut closer to the shade. The paper folds open gradually, revealing a warm amber
light inside. Keep the brass base stationary and the paper texture recognizable.
Shot 3: Cut back to a medium shot of the open lamp. A gentle push-in ends on the illuminated
shade, with space around the product for a title added later in editing.
Sound: a soft paper rustle as the shade opens, then quiet room tone.
No hands, spoken dialogue, subtitles, music, or generated lettering.The reveal supplies the motion; the lighting supports it. If the shade deforms, compare the attached image with the result and clarify the fold sequence. Asking for more “premium realism” does not tell the model which shape to preserve. Add important branding or exact typography in postproduction when you need reliable final artwork.
Place Dialogue, Voices, and Sound in the Scene
For dialogue, name the speaker before the line and state which reference supplies the voice. With multiple speakers, make the order and attribution explicit. A generic “two characters talk” leaves both voice assignment and timing unresolved.
The current tutorial documents bracket conventions for music, sound effects, dialogue, and subtitles. You can use them to distinguish instructions, but attribution and readable sentences still matter. Negative directions such as “No subtitles” and “No background music” are supported guidance, not guaranteed suppression. See the tutorial's audio tips and the prompt-guide FAQ.
A two-mascot exchange for 2.5
Attach a stylized owl image as Image1, a stylized fox image as Image2, and a permitted voice reference as Audio1. In this example Audio1 belongs only to the owl; the fox's voice is described in text. Request a 10-second clip with audio generation enabled in the interface you use.
Create a playful stop-motion scene in a tiny observatory at night.
Use Image1 for Owl, the blue felt owl with round spectacles.
Use Image2 for Fox, the orange felt fox wearing a green scarf.
Use Audio1 only for Owl's voice. Fox has a warm, low English voice.
Keep each character's voice assigned to that character throughout.
0–3 seconds: Medium two-shot. Owl turns the telescope toward the window.
<One gentle click from the telescope mount. Quiet nighttime ambience.>
3–6 seconds: Cut to Owl. Owl looks at Fox and says in English:
{Owl: "I found our missing star."}
6–10 seconds: Cut to Fox beside the telescope. Fox smiles and replies in English:
{Fox: "Then let's bring it home."}
Speak only the two quoted lines, once each. The other character listens silently
while a line is spoken. No subtitles or background music.
Preserve the felt texture, spectacles, and green scarf across shots.Keep the lines short enough for their intended windows, and review the actual speech. A requested time window does not prove a line will land at an exact second. If the line overruns, shorten the sentence or give the speaker more time before changing the visual story.
If a voice swaps characters, first check attachment order and the voice assignment at the top. If unexpected subtitles appear, inspect the input references for existing subtitles as well as the prompt's text instructions. A source reference can contribute text or sound you did not intend to copy; define its job narrowly.
Repair Matrix: Fix One Failure at a Time
Describe the observed difference between your prompt and the output before changing anything. “It looks wrong” does not identify a repair. “The acorn disappears during the carry” does.

| What you observed | What to inspect first | A focused change to try |
|---|---|---|
| Subject design changes after a cut | Reference order, competing designs, subject attribution | Identify the exact subject in Image1 and repeat the traits that changed |
| A prop vanishes or teleports | Where the shot begins and ends | State which hand or position carries the prop into the next shot |
| The action is skipped | Number of events and time available | Remove a competing beat or allocate more time to the missing action |
| Camera movement overwhelms the action | Conflicting movements in that shot | Replace the move with a fixed view or one described path |
| A voice belongs to the wrong character | Speaker names and AudioN mapping | Assign the voice to one named speaker and make the listener silent |
| Unwanted music or subtitles appear | Prompt instructions and reference content | Remove conflicting instructions and explicitly restrict the reference's sound/text role |
These are diagnostic starting points, not proven causes for every output. Choose the row that matches the result you saw, then keep the other accepted instructions stable.
Rewrite a prompt when you need a new generation
A regeneration starts another generation from your prompt and inputs. For the squirrel's missing acorn, replace the carry instruction in the complete original prompt with:
3–7 seconds: Side view at ground level. The squirrel carries the same glowing amber
acorn in both front paws for the entire walk across the moss. Keep the acorn visible
in front of its chest until it arrives at the hollow. Track alongside the squirrel.Retain the other shots and reference assignments. This is an instruction for a new result; it does not preserve the old clip's pixels simply because you write “keep everything else.”
Edit the accepted clip when the change is local
In 2.5, editing can use the existing video as Video1. The documented native edit configuration requires a 4–30-second reference video, ratio: adaptive, and duration: -1. If your interface exposes omni_reference_task_type, the corresponding value is edit. These settings support the request; they do not override its meaning. The service can still report TaskTypeMismatch when the inferred intent conflicts with the selected task. See the current task-type rules.
For a permitted 12-second squirrel clip whose acorn changes color only in the middle, use:
Edit Video1. From 3 to 7 seconds, change the acorn's blue glow to warm amber,
matching its glow at the beginning of the clip.
Preserve the squirrel's movement, camera path, mossy forest, shot order, and sound.
Keep the sections before 3 seconds and after 7 seconds unchanged.This clearly names the interval and the A-to-B change. It is a preservation request, not a promise of exact unchanged frames. Review both the repaired interval and the accepted sections, including sound. Documentation describes approximate edit-duration differences, so compare the returned clip's actual length before replacing it on an editing timeline.
Extend the ending when the story needs another beat
Use an extension when the existing clip ends correctly and you want to continue from there. For native 2.5 extension, use a reference video, ratio: adaptive, and a documented duration of 4–30 seconds or -1; the explicit task-type value, when exposed, is extend. Unlike editing, a numeric extension duration is allowed. The tutorial lists these conditions.
An extension prompt for the same story can be:
Continue Video1 from its ending inside the hollow tree.
The squirrel stays beside the amber acorn. Its glow slowly illuminates tiny copper
gears in the tree wall, and the squirrel looks up as they begin to turn.
Continue the existing camera position and dusk lighting smoothly.
Sound: the quiet forest ambience continues, with a gentle ticking from the gears.
No dialogue, subtitles, or background music.Keep the ending's objects and positions in your description. If the first extension frame introduces a new setting, the continuation is no longer doing the job you requested. Review that handoff before adding another story beat.
Decide What to Keep Before Requesting the Final Clip
Judge the video against your plan: can a viewer follow the action, do the assigned subjects and props remain recognizable, and does the sound belong to the intended moment? Keep successful creative choices. Change the part that fails your actual purpose, rather than restarting because another style adjective might look impressive.
For ModelArk 2.5, native Draft mode is a specific workflow: create a 480p draft with draft: true, then use a successful draft task ID to request a 1080p final with the same model. The final inherits the prompt, references, duration, ratio, seed, audio-generation setting, and task type; those inherited fields must not be resent, even unchanged. The draft ID is usable for seven days, and both steps are billed separately. The Draft-mode section documents the procedure.
Use that workflow when you want to finalize a successful native draft. If you need to change the prompt or references, create a new draft or generation instead. A third-party “preview” button may follow a different process, and Draft mode is not a free preview or a promise of identical final pixels.
Save the prompt, reference order, requested settings, and the result you accepted together. That record makes the next repair understandable. If the job never produced a video because it was rejected, stalled, or timed out, move to accepted-task troubleshooting before rewriting the scene. A generation failure and an unsatisfactory generated clip require different next actions.
Frequently Asked Questions
What is the best Seedance 2 prompt structure?
Start with the visible action and ending, assign references to specific subjects or jobs, and describe each shot's framing, movement, and sound. Use continuity instructions for details that must persist. The prompt card above is a reusable planning tool; it is not a required syntax or a tested guarantee of better clips.
Can I use timestamps in Seedance 2 prompts?
Seedance 2.5 supports integer-second timestamps as well as Shot N. For 2.0, use shot labels; a written timing budget can guide your planning but does not give it 2.5's timestamp control. Even in 2.5, timecodes do not guarantee frame-exact cuts. Source: current prompt guide.
How many reference images should I use?
Use the images that have distinct jobs. Start with the subject and setting you need, then add another view or a movement reference when it resolves a specific uncertainty. Native ModelArk 2.5 permits up to 30 images within 50 total reference assets; the 2.0 family permits up to nine images within 15 total. Those are maximums, not recommended counts, and your app may impose different limits. Source: version comparison and input rules.
Why does my app cut off or reject a long Seedance prompt?
No documented model limit explains it, so the cap most likely comes from the app or service you use. As of November 8, 2026, neither the BytePlus ModelArk nor the Volcengine Ark documentation publishes a hard character limit for Seedance prompts, but the apps do: Runway allows 10,000 characters for Seedance 2.0 and 15,000 for 2.5, and invideo allows 3,000. Check the help pages of the interface you use for its own box limit. Shortening the prompt is usually the better fix anyway, because ByteDance's own guidance warns that overly long prompts make the model drop details.
Does calling an image “the first frame” lock the starting image?
A text instruction alone does not give a semantic reference image the native first_frame role. Use the actual frame input mode or role your interface provides. For ModelArk 2.5, that role requires ratio: adaptive; it constrains the start differently from general visual referencing. Source: reference and frame modes.
How do I fix identity drift without changing the whole scene?
Check which attachment supplies the subject and which part of it you mean. Clarify that mapping, remove a conflicting design, and repeat the specific features that drifted. If an existing 2.5 clip is otherwise accepted, consider a focused edit rather than assuming a new generation will preserve it. Check the untouched sections after the edit, too.
Can “No subtitles” or “No music” guarantee their removal?
No. These are supported prompt directions, but the official FAQ says suppression can still fail. Check whether a reference includes the unwanted element and explicitly limit what that reference should supply. For a deliverable requiring clean audio or exact text placement, review the result and finish it in your editor. Source: prompt-guide FAQ.
Is this also a Seedance 2 API guide?
This workbook covers the creative input and the task settings that change its meaning. For current native model IDs, request fields, and the Draft-to-final procedure, use the official ModelArk tutorial. Authentication, provider selection, and billing setup are separate tasks; writing a scene prompt does not establish access to a model.
Sources7
External pages this guide links to, in the order they appear. Last updated Nov 8, 2026.
Sources7
External pages this guide links to, in the order they appear. Last updated Nov 8, 2026.
- 1.current BytePlus prompt guidedocs.byteplus.com/en/docs/modelark/seedance-2-5-prompt-guide
- 2.2.5 tutorial and model tabledocs.byteplus.com/en/docs/modelark/seedance-2-5
- 3.Volcengine Ark video generation API referencedocs.volcengine.com/docs/ark/create-video-generation-task-api
- 4.Runway Help Center, "Creating with Seedance 2.0"help.runwayml.com/hc/en-us/articles/50488490233363-Creating-with-Seedance-2-0
- 5.Runway Help Center, "Creating with Seedance 2.5"help.runwayml.com/hc/en-us/articles/53542207042323
- 6.invideo's Seedance 2.0 prompt guideinvideo.io/blog/seedance-2-0-prompt-guide
- 7.See the portrait-asset guidedocs.byteplus.com/en/docs/modelark/seedance-portrait-asset-guide





