Seedance Prompting: How to Prompt Seedance 2.5 and 2.0

TL;DR: Seedance prompting is plain prose plus three documented rules: bind assets as @Image 1 / @Video 1 / @Audio 1, sequence events with Shot 1 / Shot 2, and mark sound with (), <>, {} and 【】. Seedance 2.5 arrived on 31 July 2026 with 30-second output and second-level timestamps. Seedance 2.0 ignores timestamps entirely.

Which Seedance are you actually prompting?

Seedance 2.5, unless you pinned a model ID. ByteDance Seed published the launch post One-take Creation, Flexible Referencing: Introducing Seedance 2.5 dated 31 July 2026, describing 30-second single-pass generation, up to 30 images plus 10 video clips plus 10 audio clips as references, and timestamp-level editing. At that point it said API access was "coming soon via BytePlus ModelArk". It has since arrived: the create-task API reference now opens with a banner stating that Dreamina Seedance 2.5 is fully available and can be called on the platform (read 27 August 2026).

This matters more than a version bump usually does, because the prompting rules are not the same across the two. Most guides on this query, including pages that look current, describe 2.0 behaviour and present it as how Seedance works.

Seedance 2.0 has not gone anywhere. Its launch post is dated 12 February 2026, and 2.0, 2.0 fast and 2.0 mini are still separately documented, separately priced and callable today. It also remains the only member of the family that outputs 4K. This is not a migration story, it is a "know which one you are talking to" story.

What has actually changed between Seedance 2.0 and 2.5?

ByteDance publishes the answer itself, in a section of the Seedance 2.5 prompt guide headed "Differences from Seedance 2.0". Four items, quoted from that page (read 27 August 2026):

1. "Seedance 2.0 does not respond to timestamps and only responds to shot numbers, while Seedance 2.5 supports integer-second timestamps."
2. "Seedance 2.0 does not recommend using multi-view images as subject references, while Seedance 2.5 supports them."
3. "Seedance 2.0 only supports six fixed output aspect ratios, while Seedance 2.5 can support any output aspect ratio between [0.4, 2.5] by controlling the input assets."
4. "Seedance 2.5 supports MOV output, which better preserves color consistency, brightness consistency, and audio-visual consistency in extension and editing tasks."

Item one is the one that silently ruins prompts. Everything else follows from the capability table:

Note the asymmetry that trips people up: 4K is a 2.0 privilege, not a 2.5 one, and it is restricted to the base 2.0 model. Fast and mini top out at 720p.

What is the Seedance prompt formula?

ByteDance gives a different formula per version, and both are short enough to memorise.

The Seedance 2.5 tutorial orders it as subject, then action or event, then scene and environment, then visual style, then camera movement and cuts, then sound. Drop anything you do not need.

The 2.0 series prompt guide prints a longer "advanced formula", and the reasoning behind it is the useful part. It describes the model as decomposing your input into a "spatial layer" of what is in the frame and a "temporal layer" of how it changes, so a good prompt is "not simply copywriting-style description, but an engineering-style instruction":

precise subject + action details + scene/environment + lighting & color tone
+ camera movement + visual style + image quality + constraints

Both guides recommend the same overall shape: one sentence of summary, then a segmented body, then a closing paragraph of details that persist across the whole clip. The 2.5 guide caps it with a length recommendation that applies across the family: "no more than 500 Chinese characters or 1,000 English words", because longer text scatters the model's attention and elements go missing.

If you want the general theory rather than the vendor-specific version, our breakdown of the seven parts of a video prompt covers the same ground model-agnostically.

How do you reference images, video and audio in a Seedance prompt?

By type and ordinal position. You count from 1 within each asset type, in the order the assets appear in your request body. ByteDance writes them as Image 1, Video 1, Audio 1, usually with an @ prefix in prompt text.

The rule is stated most bluntly in the portrait-video doc: "Prompts must reference assets in the format asset type + number, where the number is the sorting order of the asset among assets of the same type in the request body." The same page shows the failure it prevents, using the asset library:

Correct:   The beauty influencer in Image 1
Incorrect: asset-2026**** is a beauty influencer

The model cannot associate an opaque asset ID with anything semantic, so an ID pasted into prompt text is dead weight.

For people and objects that recur, bind a label once and reuse it. The 2.0 prompt guide's pattern is to name two or three stable static features:

Define the woman wearing a red dress and a straw hat in Image 1 as Subject 1.
Define the tall man in Video 1 as the police officer, and the shorter man as the thief.

For one-off subjects that were never defined, the guide supports an inline binding written as Subject N@Image N, for example Zhang San@Image 1. ByteDance's own Seedance 2.0 prompt-optimizer skill, published as plain markdown at arkdocs.tos-cn-beijing.volces.com, adds a parsing caution worth stealing: do not let @Image 1 run straight into a verb, because the digit and the following word can fuse ambiguously. Write Subject 1@Image 1 runs, or insert a noun, as in the woman in Image 1 runs.

Be specific about what each asset supplies. Three patterns from the 2.5 guide:

Images 1-2 are Character 1 and correspond to Audio 1; Images 3-4 are Character 2 and correspond to Audio 2.
Refer to the action of casting the spell in Video 1 and the wrap-around camera movement in Video 2.
Strictly refer to the actions and camera movements in Video 1, and keep the sequence consistent with the video.

That third one carries an instruction most people miss: when the reference is already precise, stop describing it in words. The guide says there is "no need to describe details such as raising a hand, turning around, or having the camera slowly orbit". Redundant prose fights the asset.

Shot numbers or timestamps: which does Seedance understand?

Both, on 2.5. Only shot numbers, on 2.0.

The 2.0 prompt guide is unambiguous: "Use identifiers such as Shot 1, Shot 2, and Shot 3 to organize content in the order events occur", and, in a tip immediately below, "The model's support for precise timing (such as 0–3 seconds) is unstable, and forcibly limiting duration may lead to abnormal generation results." ByteDance's own 2.0 optimizer skill states the same rule as a hard constraint, forbidding absolute seconds and telling you to prefer shot order.

Seedance 2.5 flips this. Its prompt guide says either format is acceptable, then documents three timing mechanisms in 1-second units: explicit intervals, single time points, and relative timing.

Shot 1: Low-angle side shot of an alley. The man starts running, breathing hard.
Shot 2: He knocks over a fruit stand. The camera shakes and cuts to a close-up of his frightened face.
Shot 3: He climbs a low wall and disappears. The camera pulls back and holds on the empty street.

0-3s: wide shot of the empty platform, the train's headlight growing in the tunnel mouth.
3-7s: the doors open, the boy bolts out, the camera whip-pans to follow.
7-15s: the man catches him on the concourse; the camera settles into a handheld two-shot.

Two cautions on the 2.5 timeline. Keep it continuous, because the guide warns against gaps such as "0-3s... 5-6s...". And do not use timestamps for frequency: it says not to write things like "shake your head three times per second".

One more rule that survives both versions, from the optimizer skill: one camera move per shot. Push, pull, pan, track, orbit or locked off, pick a single one, and do not stack them. If you need the vocabulary for that, our camera movement reference lists the terms these models recognise.

How do you control audio in a Seedance prompt?

Audio is on by default. generate_audio defaults to true on Seedance 2.5, the 2.0 series and 1.5 pro, so a plain text prompt returns a video with speech, effects and music unless you switch it off. The API reference also recommends putting dialogue inside double quotes for better audio results.

On top of that, the 2.0 prompt guide publishes a character convention, and the 2.5 tutorial repeats it:

Dialogue in a language other than Chinese or English must be tagged, as in says in Japanese {こんにちは}. Keep one language per line; mixing is called out as a failure mode, proper nouns excepted.

Negation is where Seedance differs from most video models. It has no negative_prompt parameter at all, and the 2.5 guide restricts negative phrasing to two narrow categories: subtitles and audio. "No subtitles", "No BGM; generate only environmental sounds and action sounds", and "No audio" are all documented as supported. Anything else should be phrased positively. That is consistent with what we found building our negative prompt support matrix, where Seedance sits in the "no parameter published" column.

Can you use a real person's face as a Seedance reference?

No, and this kills the workflow most people reach for first. The API reference states that "Dreamina Seedance 2.5 and Dreamina Seedance 2.0 series models do not support directly uploading reference images or videos that contain real human faces."

Three sanctioned routes exist instead, documented on ModelArk's portrait-video page:

• Trusted model output. Face-bearing video your own account generated with Seedance 2.5 or the 2.0 series can be fed back in as a reference for 30 days from generation. Same account, same platform only. Last-frame images from those videos and Seedream 5.0 lite text-to-image outputs are covered on the same 30-day terms.
• Preset digital characters. A platform library of compliant portrait assets, passed as asset://<asset ID> and referenced in the prompt as Image 1 like anything else.
• Authorised real-person assets, for material you hold rights to.

The practical consequence: build your cast inside the platform, in the same account, and start the clock. A character you generated 31 days ago is no longer usable as a reference.

Which parameters actually apply to Seedance 2.x?

This is the trap the flag syntax sets. ModelArk documents that resolution, ratio, duration, frames, seed, camera_fixed and watermark can all be appended to the prompt text as --[parameters] instead of being sent in the request body, and says all models accept that mechanism. Separately, it says supported parameters vary by model, and that appended flags use weak validation: "Invalid parameters are ignored or cause an error." Body parameters are validated strictly and return a real error.

Put those together and you get a silent failure. seed, camera_fixed and frames are each documented with a supported-models list containing only Seedance 1.5 pro, 1.0 pro and 1.0 pro fast. Append --cf true --seed 42 to a Seedance 2.0 prompt and nothing errors, nothing changes, and you spend an afternoon deciding the model is inconsistent.

{
  "model": "dreamina-seedance-2-5-260628",
  "content": [
    { "type": "text", "text": "Shot 1: ..." },
    { "type": "image_url", "image_url": { "url": "https://..." }, "role": "reference_image" },
    { "type": "video_url", "video_url": { "url": "https://..." }, "role": "reference_video" }
  ],
  "generate_audio": true,
  "ratio": "adaptive",
  "duration": -1,
  "omni_reference_task_type": "reference",
  "output_format": "mov"
}

The trigger words matter, because that is how 2.5 classifies the job. The tutorial lists them: an edit needs one of edit / add / delete / modify / replace; an extension needs extend forward or backward / continue. Which brings us to the templates.

Copy-paste Seedance prompts

Eighteen patterns, written in the syntax the docs use. Swap the nouns.

Single subject, 2.0 series, shot-numbered

Define the woman in the red jacket in Image 1 as Subject 1.
Shot 1: Medium shot, camera locked off. Subject 1 steps into a rain-soaked doorway and shakes water from her sleeve.
Shot 2: Slow push in to a close-up. She looks off-frame left, shoulders dropping, and exhales.
Shot 3: Camera pulls back to a wide. She walks out of frame right; the doorway stays empty for a beat.
Overall: cold blue-grey grade, cinematic realism, soft natural light. Faces stable and undistorted,
movement continuous and natural. Keep it subtitle-free. Do not generate watermarks or logos.

Same scene, 2.5, timestamped

0-4s: Medium shot, camera locked off. The woman from Image 1 steps into a rain-soaked doorway and shakes water from her sleeve.
4-9s: Slow push in to a close-up. She looks off-frame left, shoulders dropping, and exhales.
9-15s: The camera pulls back to a wide. She walks out of frame right; the doorway holds empty.
Overall: cold blue-grey grade, cinematic realism, soft natural light, one continuous take. No subtitles.

Two subjects, explicit mapping

Images 1-2 are Character 1 and correspond to Audio 1; Images 3-4 are Character 2 and correspond to Audio 2.
Character 1's facial features reference Image 1 (headshot); makeup and styling reference Image 2 (full-body).
Shot 1: Wide two-shot in a daylight market. Character 1 stands left of frame, Character 2 right of frame.
Shot 2: Handheld tracking shot. Character 2 pushes through the stalls; Character 1 follows.
Throughout, characters with identical appearance, clothing and accessories are prohibited. Do not generate twins
or duplicate copies of a character in the same frame.

Partial reference, motion only

Refer to Video 1 for the orbiting camera movement only, and Image 1 for the subject's appearance and costume.
Do not take lighting or background from Video 1.
Shot 1: The subject raises both arms as the camera completes one slow orbit at chest height.

Strict motion transfer

Strictly refer to the actions and camera movements in Video 1, and keep the sequence consistent with the video.
Replace the performer with the subject in Image 1. Keep the stage, lighting and audience unchanged.

First and last frame, named in prompt

Image 3 is the first frame, and Image 5 is the last frame.
Between them, the camera drifts slowly left while the fog thickens and the streetlights come on one by one.

Keyframe order

Use Images 1 through 7 in order as keyframes. A blue-and-pink long-tailed fish glides above a sea of cloud,
the camera pushes toward a mountain town, then cuts inside a wood-panelled hall where the fish circles a stone pool.
Ukiyo-e-influenced illustration style, continuous camera logic between keyframes.

Multi-panel storyboard, 2.5

Asset binding: storyboard panels @Image 1, bedroom @Image 2, Li Tian @Image 3, Li Qian @Image 4.
Follow the panel order to build one continuous story with coherent camera logic.
Shot 1: [wide, locked off, rule-of-thirds] Snow falls past a floor-to-ceiling window; the man stands with his back to camera.
Shot 2: [medium over-shoulder] He turns and looks at her; she lowers her head.
Shot 3: [close-up, centred] She holds a book against her chest, eyes reddening.

Clay render reference, 2.5

Refer to Clay Render 1 for camera movement, pacing, shot-size transitions, subject trajectory and blocking.
Refer to Image 2 for character design, materials, lighting, colour and atmosphere.
Render it as a warm 3D animated short with a storybook feel. Do not add shots and do not change the shot order.

Voice timbre reference

Refer to the timbre in Audio 1 for the narrator's voice: a low, warm, slightly grainy middle-aged male voice.
Keep the delivery style close to the reference. He says {Some doors you only get to walk through once}.

Full audio block, all four conventions

Shot 1: Interior night, kitchen. (a slow lo-fi piano loop plays under the scene)
<a kettle clicks off and settles> The man leans on the counter and says {You're up late}.
Shot 2: She sits down opposite him. 【Chapter Two: The Long Way Home】 appears briefly, lower third.

Non-English dialogue

Shot 1: Close-up of the shopkeeper counting coins. She looks up and says in Japanese {いらっしゃいませ}.
Shot 2: The customer nods and replies in Japanese {ありがとうございます}. Keep dialogue in one language throughout.

Silent output

Shot 1: A drone rises over a fog-filled valley at dawn, revealing a single lit farmhouse.
No BGM; generate only environmental sounds and action sounds. No subtitles.

Edit, scoped by timestamp, 2.5

Edit Video 1. Change the man's action from drinking coffee to mopping the floor from 4-6 seconds,
and leave the rest of the content unchanged.

Edit, swap a subject

Editing task: replace the woman on the right in Video 1 with the woman from Image 1.
Keep her position, timing, wardrobe silhouette and the camera move exactly as they are.

Edit, audio only

Edit Video 1. Delete the background music and keep the dialogue and ambient sound unchanged.
Do not alter the picture.

Extend backward

Extend Video 1 backward. Continue the visuals and subjects from Video 1 for another 15 seconds,
keeping the characters, scene, visual style and sound effects consistent.
The boy runs along the platform, the man catches up, and the two walk out together.

Green-screen replacement, 2.5

Using Video 1, render the green-screen background, obstacles and supporting characters.
0-4s: outdoor training; replace the obstacles with rocks, bricks, tyres and wooden crates.
4-10s: locker room, teammates offering encouragement.
10-15s: a stadium match; replace the training poles with defenders and a goalkeeper, and the protagonist scores.
Photorealistic, cinematic quality throughout.

What breaks Seedance prompts most often?

The 2.0 prompt guide ships a FAQ that reads like a defect log, and its causes carry over to 2.5.

Character ID drift. A face changes mid-clip, or drifts toward a celebrity and gets blocked at review. ByteDance's diagnosis is that the face reference is too weak, buried in a composite image where the face occupies too little area. The fix is a separate headshot with no expression and minimal background, paired with a full-body shot, plus placing the most precision-critical asset earliest in the prompt.

Duplicate characters. Two identical people in one frame, caused by undefined subjects and, especially, by multi-view character sheets, which 2.0 tends to read as several different people. The guide is explicit that this cannot be prevented outright, only made less likely, via per-character mapping and a global constraint sentence. Seedance 2.5 lifts the multi-view restriction, per its own differences list.

More than four reference people. Stability drops and headcounts go wrong. The documented workaround is to generate grouped images of at most four people each, then use those as references for the video.

Unwanted subtitles. Also documented as unpreventable, only reducible. Three levers: an explicit constraint in the prompt, stripping text out of reference images first, and preferring landscape output, where the guide says subtitles appear significantly less often than in portrait.

Clipped audio at the end. Narrated clips can end on an abrupt click. ByteDance suggests regenerating or fading the tail in an editor.

If your clip is warping rather than misbehaving in these specific ways, that is usually a different problem, and we covered it separately in why AI video morphs and warps.

Where can you actually run Seedance, and what does it cost?

Access is the part most guides skip. Both clouds gate the Seedance 2.x family behind a spend threshold, and neither has a free tier for it.

• Volcano Engine Ark (China, ark.cn-beijing.volces.com) requires an account balance above 200 RMB, a savings plan at the 200 RMB tier or above, or a purchased Seedance resource pack with quota remaining.
• BytePlus ModelArk (international, ark.ap-southeast.bytepluses.com) requires the same three options at USD 30.

Promotional pricing was running at the time of writing, published on the create-task API reference and read on 27 August 2026: Seedance 2.5 at 1080p billed at 72% of list from 14 August to 17 September 2026, from around USD 0.41 per second; 2.0 mini at 40% of list from 7 August to 7 September, from about USD 0.03 per second at 720p; 2.0 fast at 75% over the same window, from about USD 0.09 per second at 720p. Those windows close, so check the page, not this paragraph.

Two operational limits from the Seedance 2.5 tutorial are easy to miss: task records are kept for 7 days, and the generated video URL expires after 24 hours with a cap of 100 downloads. Pull your files down promptly.

Treat the leaderboards carefully. On Artificial Analysis's Video Arena text-to-video boards, read on 27 August 2026, Dreamina Seedance 2.0 720p sits inside the top five with audio and without it, rather than at the top, and Seedance 2.5 was not listed at all. Our Seedance 2.0 prompt templates post goes deeper on 2.0 syntax if that is the version you are pinned to.

Where a prompt tool fits, and where it does not

Prompt Architects does not generate video. It generates the prompt, then keeps it. That distinction matters here, because a Seedance prompt is a structured artefact: a subject-binding block, a shot list, and a constraints tail that is identical across every clip you ever make. That is a prompt template with variables, not a piece of writing you retype.

Video prompt generation sits on the Advanced and Team plans; the video prompt library is on every tier, including the free one, which the FAQ page describes as 5 prompt enhancements per day, forever. If you are shooting a series in Seedance, the win is not a cleverer adjective. It is that shot four of episode nine still carries the same twin-prevention sentence as shot one of episode one.