Seedance 2.0 Prompt Templates: The Syntax ByteDance Documents

TL;DR: Seedance 2.0 prompts are ordinary prose plus three rules ByteDance actually documents: reference assets as @Image 1 / @Video 1 / @Audio 1, bind people with Subject 1@Image 1, and sequence events as Shot 1 / Shot 2 rather than timestamps. Duration runs 4 to 15 seconds at a fixed 24fps, and audio is on by default.

What is the Seedance 2.0 prompt format?

Seedance 2.0 takes natural language, not a parameter string. The structure that its own documentation asks for is a subject definition, then a numbered shot list, then a block of style and constraint phrases at the end.

That last part is the piece most guides miss. ByteDance publishes a prompt-optimizer skill file for the model, hosted on its own object storage at arkdocs.tos-cn-beijing.volces.com (retrieved August 26, 2026), and it is blunt about the philosophy: a good Seedance prompt is "an engineering instruction, not a descriptive blurb." The model splits your input into a spatial layer (what is in the frame) and a temporal layer (how it changes), and the prompt is expected to feed both.

There is a second, optional layer. Volcano Engine Ark's create-task API reference documents that resolution, ratio, duration, frames, seed, camera_fixed and watermark can be appended to the prompt text as flags instead of being passed in the request body:

A kitten yawns at the camera --rs 720p --rt 16:9 --dur 5 --seed 11 --wm true

The doc labels this "weak validation" and warns that a malformed flag is either ignored or throws. Request-body parameters are strongly validated. Use flags when you are pasting into a playground, use the body when you are writing code. Source: Volcano Engine Ark, "创建视频生成任务", doc updated August 21, 2026, retrieved August 26, 2026.

Which Seedance model are you actually prompting?

There are three models in the 2.0 family and they share one prompt syntax. Ark's tutorial states that their capabilities are basically identical and that the difference is quality against cost.

One thing to get straight before you invest a weekend in this syntax. Seedance 2.0 is no longer the newest model in the line. Ark documents Seedance 2.5 as doubao-seedance-2-5-260628, with 30-second single-call generation, up to 50 reference assets, and timestamp-based editing control (doc 2607688, first published August 7, 2026). ByteDance Seed's own Models navigation lists Seedance 2.5, not 2.0.

The syntax carries over almost entirely, so learning it on 2.0 is not wasted. But 2.0 keeps one advantage: it is the only member of the family Ark lists as supporting 4K output, in 10-bit depth with H.265/HEVC encoding.

Can you actually get access to Seedance 2.0?

This is the part the reseller blogs skip, and it is the part that decides whether the rest of this page is useful to you.

Volcano Engine Ark serves the model from the cn-beijing region. Before you can enable it, the tutorial requires one of the following: an account balance above 200 RMB, a savings plan purchased at the 200 RMB tier or higher, or an existing Seedance resource pack with quota remaining. There is no free tier and no trial credit path documented on that page.

BytePlus ModelArk is ByteDance's international platform and carries the same documentation set in English, under the product name Dreamina Seedance 2.0. The English pages exist at the same document IDs, so docs.byteplus.com/en/docs/ModelArk/2291680 is the English twin of the Chinese tutorial. Both render only with JavaScript. A plain fetch of either returns navigation and no body text, which is a large part of why so little accurate Seedance prompt guidance exists in English.

There is a harder constraint than billing. Ark's API reference states that neither the Seedance 2.0 series nor Seedance 2.5 accepts reference images or videos containing real human faces. Volcano Engine offers two sanctioned alternatives: a built-in library of virtual portraits, and a separate flow for portrait material you already hold rights to. If your plan was to feed the model a photo of a client or an actor and get character consistency for free, it will not work that way.

What syntax does ByteDance actually document?

Four conventions, all from ByteDance-owned pages.

1. Assets are referenced by type and index. Ark's tutorial states that prompts must reference material using a "type plus number" format, where the number is that asset's position among assets of the same type in the request body, counting from 1. It adds explicitly that Asset IDs cannot be used to refer to material inside the prompt text. ByteDance's prompt-optimizer skill standardises this as @Image 1, @Video 1, @Audio 1. The English launch post uses the same shape, with a sample prompt reading "Refer to the shooting script in @Image 1 ... The character is from @Image 2, the scene is from @Image 3."

2. Subjects get bound to their source. Write Subject 1@Image 1 on first mention, or define the binding once at the top and reuse a stable label thereafter. The skill file warns against writing @Image 1 immediately before a verb, because the digits run into the following word and create ambiguity. Put a noun in between.

3. Time is expressed as shots, never seconds. Both the prompt guide and the optimizer skill prohibit absolute timings such as "0 to 3s" and state that the model's support for precise timing is unstable. Use Shot 1 / Shot 2 / Shot 3 and let the model pace it. Each shot is organised in a fixed order: camera move, then subject action and expression, then position change, then audio.

4. Different information types get different brackets. This is the single most useful undocumented-in-English detail in the whole skill file:

Worth flagging a small inconsistency inside ByteDance's own documentation. The API reference for generate_audio recommends placing dialogue inside double quotation marks to improve audio generation, while the optimizer skill specifies curly braces. Both are ByteDance sources, both current as of August 2026. Pick one and stay consistent within a prompt.

What are the real duration, resolution and audio limits?

Every number below comes from Ark's create-task API reference, doc 2222480's sibling doc 1520757, updated August 21, 2026 and retrieved August 26, 2026.

Two of those deserve a second look.

The reference budget is smaller than it sounds. Nine images, three videos and three audio clips is the ceiling, but each reference video must be 2 to 15 seconds and all videos together must total no more than 15 seconds. Same arithmetic for audio. And the prompt guide advises against filling the budget at all: it recommends four to five assets total, one or two character images plus one scene image plus one camera-movement video plus one audio clip, on the grounds that too many references make the model unable to rank feature priority.

The audio channel count is contested. ByteDance Seed's launch post from February 12, 2026 describes "dual-channel audio" and "two-channel stereo technology" as a headline feature (seed.bytedance.com, retrieved August 26, 2026). Ark's API reference, updated six months later, states that generated audio is always mono irrespective of how many channels the input audio had. I could not reconcile these from published sources. If stereo output matters to your pipeline, test it rather than trusting either page.

The three task types, and the phrasing that triggers each

Seedance 2.0 infers what you want from your wording, and getting the wording wrong is the most common way to waste a generation. Volcano Engine's prompt guide splits reference-driven work into three tasks with three sentence patterns.

# 1. MULTIMODAL REFERENCE: pull elements out, make something new
Refer to <Subject 1> in @Image 1, and generate ...
Refer to the <camera movement / motion / style / sound design> in @Video 1, and generate ...
Refer to the vocal timbre in @Audio 1, and generate ...

# 2. EDIT: change part of an existing video, leave the rest alone
Strictly edit @Video 1, changing <original feature> to <new feature>.

# 3. EXTEND: continue an existing video forwards or backwards
Extend @Video 1 forwards / backwards, generating ...
Track completion: @Video 1, <transition description>, then @Video 2

Copy-paste Seedance prompt templates

Two structures cover nearly everything. ByteDance's optimizer skill calls them Path A and Path B, and the split is by event density rather than asset count: one continuous action in one place is Path A, anything with multiple events or multiple locations is Path B.

Template 1: simple reference (Path A). One paragraph, no headings, constraints folded into the last sentence.

Refer to <Subject 1> in @Image 1 (short-haired woman), and generate her sitting
by the window of the cafe in @Image 2, eating cake with quiet concentration,
warm yellow light falling softly across the table.
High-definition cinematic texture, stable image with no deformation,
keep it subtitle-free, do not generate a watermark, do not generate a logo.

Template 2: multi-shot narrative (Path B). Three blocks, in this order, every time.

[BLOCK 1: SETUP AND SUBJECT BINDING]
Overall setting: a modern office in late afternoon, natural soft lighting,
understated documentary feel.
<Subject 1>'s facial features reference @Image 1 (head shot); wardrobe and
styling reference @Image 2 (full-length shot).
Define the minimal wooden meeting room in @Image 3 as <Scene 1>.
Camera language references the mid-shot push-in and gentle pan of @Video 1.
Ambient tone references @Audio 1.

[BLOCK 2: SHOT LIST]
Shot 1: Steady mid-shot tracking. <Subject 1> walks to the door of <Scene 1>
at an even pace, pauses, takes one slow breath, expression slightly tense.
<footsteps on carpet, muffled conversation further down the hall>

Shot 2: Cut to an interior mid-shot. <Subject 1> pushes the door open. Two
colleagues look up from their laptops. One smiles and asks {So? How did it go?}
The camera moves slowly between them in half-body close-ups.

Shot 3: Close-up. <Subject 1> looks down, shoulders dropping, then lifts her
head and breaks into a grin: {I got it.} The camera pulls slowly back and
settles on a wide of the room.
（light acoustic guitar rises under the final line）

[BLOCK 3: STYLE AND CONSTRAINTS]
High-definition cinematic documentary style, warm palette, soft light and shadow.
Faces stable and undistorted, features clear, movement continuous and natural,
never stiff, no clipping, no stutter.
Keep it subtitle-free, avoid generating any text or subtitles.
Do not generate a watermark. Do not generate a logo.
No two characters may share identical appearance, clothing or accessories at any
point; do not generate duplicate or twin figures.

Template 3: targeted edit.

Strictly edit @Video 1, replacing the perfume bottle on the counter with the
face cream in @Image 1. Motion and camera movement unchanged.
Stable image with no deformation, do not generate a watermark, do not generate a logo.

Template 4: extension with a controlled hand-off.

Extend @Video 1 backwards, generating the two of them continuing toward the
corner of the street and exchanging a brief smile, then ending on @Video 1.
Stable image with no deformation, keep it subtitle-free,
do not generate a watermark, do not generate a logo.

That final "then ending on @Video 1" is not decoration. The tutorial notes that a forward or backward extension normally returns only the tail of the original clip, and that you can pull the original footage back into the output by naming it at the end of the prompt.

Template 5: on-screen text. The prompt guide gives three shapes, one each for a tagline, a subtitle and a speech bubble.

Tagline:  <text content> + <when it appears> + <where it appears> +
          <how it appears>, <colour and style>
Subtitle: A subtitle appears at the bottom of the frame reading "...",
          fully synchronised with the audio rhythm
Bubble:   <Character> says: "...", and a speech bubble appears beside them
          containing the line

Stick to common characters. The guide warns that rare glyphs and unusual symbols degrade text rendering.

If you keep these five as a prompt template set rather than retyping them, the same discipline that makes JSON video prompts for Veo 3 reusable applies here: the constraint block is the part you never want to rewrite from memory.

What actually breaks Seedance prompts

Volcano Engine publishes a troubleshooting section, which is unusually candid for a model page. The failures below are theirs, not mine.

Identity drift. Faces change mid-clip when the reference image mixes a face with a full-body pose or an outfit shot, or when the face occupies too little of the frame. The fix is a dedicated head shot, face only, neutral expression, minimal background, placed early in the prompt. The guide explicitly advises against character turnaround sheets or multi-view reference images, because the model reads different angles as different people.

Twins. Same root cause. Multiple similar-looking figures in one frame. The documented mitigation is to tag each person with their source image, as in "Zhang San (per Image 1) throws the folder to Li Si (per Image 2)", and to append a global no-duplicates constraint.

Unrequested subtitles and watermarks. Both are acknowledged as impossible to eliminate completely. The guidance is to add explicit negative constraints, to strip text out of reference material first, and, where the brief allows it, to generate landscape rather than portrait, because portrait output produces subtitles noticeably more often.

Style drift. A realistic reference image plus a prompt that never states a style will drift toward photorealism even when you wanted animation. Anchor the style word explicitly.

Seams and degradation on extension. Joining an extended clip to the original can jump at the splice. The documented workaround is oddly specific: trim six frames off the end of the earlier clip and one frame off the start of the later one, then repeat at every join. Quality also degrades cumulatively across repeated extensions, so the guide advises limiting how many times you extend the same footage.

Most of these are the same failure classes that make AI video look generic on every other model. The difference is that ByteDance wrote them down.

How Seedance 2.0 compares to what else is documented

Careful here, because the video-model spec sheets circulating online are mostly wrong.

Two notes on that table. Google's Gemini API documentation recommends Omni Flash as the default model for video generation, but that recommendation is scoped to those docs; DeepMind's model card does not designate it a default. And the widely repeated "Kling 3.0 does 15 seconds at 4K and 60fps" figure does not survive checking. The 4K tier is corroborated by Kling's own pricing, but the duration and frame rate are not stated on any Kling-owned page we could reach.

On leaderboards, Seedance 2.0 is genuinely strong but not uniformly first. Artificial Analysis, checked August 26, 2026, ranks Dreamina Seedance 2.0 720p first in image-to-video with audio at 1,192 Elo, ahead of MiniMax H3 and Gemini Omni Flash. On text-to-video with audio the same model sits fourth at 1,220, behind Wan 3.0, Gemini Omni Flash and MiniMax H3. If your work starts from a still image, that first ranking is the relevant one. If you start from text, it is not.

Where a prompt tool fits, and where it does not

Plain honesty, since this is our blog. Prompt Architects does not generate video. We generate the prompt. Seedance renders it, and nothing on our side changes what comes back from Ark.

What the video prompt feature does is stop you retyping the three-block structure and the constraint stack every time, and keep the variants somewhere you can find them again. Save the constraint block as a reusable snippet, keep your subject bindings as variables, and the actual creative work becomes the shot list rather than the boilerplate. If you have ever lost a prompt that worked, that is the problem being solved, and it is the same one behind our template library.

For direction, framing and lens language that transfers across models, directing AI video like a filmmaker covers the vocabulary Seedance responds to, and the camera prompt set is a decent starting bank. Just remember the one-camera-move-per-shot rule: Seedance's own guide says stacking a push, a pan and a tilt into one shot destabilises the frame.