Seedance 2.5 prompting

Checked on 7 August 2026, the day ByteDance published the guide. Every quotation below is from ByteDance's own documentation, in English on BytePlus ModelArk or in Chinese on Volcano Engine Ark. Both docs sites render from JavaScript, so a plain fetch returns an empty page: the text lives in the window._ROUTER_DATA payload of the served HTML.

ByteDance published a prompting guide for Seedance 2.5 on 7 August 2026, a week after the model itself launched. It went up twice on the same morning: as Dreamina Seedance 2.5 prompt guide in English on BytePlus ModelArk at 05:46 UTC, and as Doubao Seedance 2.5 提示词指南 in Chinese on Volcano Engine Ark a few seconds earlier. Same document ID, 2607689, two brand names for one model. A companion page carrying the hard limits went up between them.

The guide is worth reading closely for one reason above the rest: the prompt structure it prescribes is not the structure most people are currently using.

The six-part formula going round is not ByteDance’s

A “six-part formula” has been circulating since 5 August under the heading The Official Seedance 2.5 Prompt Guide: Subject + Action or Event + Scene and Environment + Visual Style + Camera Movement or Cut + Audio. It appears in a Segmind blog post that cites no ByteDance source, and it predates ByteDance’s guide by two days.

ByteDance’s actual instruction is a four-block document. The nearest thing to a formula sits inside it as a single summary line, and its fields are different:

Treat Seedance 2.5 as a visual content producer, and write structured prompts with a visual storytelling mindset.

The Chinese edition puts it as 把 Seedance 2.5 当作一个视觉内容生产者，用导演的思维书写 [ 结构化 Prompt ]: write with a director’s thinking. Four blocks follow.

The four blocks

The structure ByteDance prescribes

Written out, the guide’s four blocks are:

1. Asset Referencing for R2V. “Clearly identify each image, video, or audio asset by its upload order and intended purpose, such as which asset represents the subject, voice, action, scene, and so on.”
2. One-Sentence Summary. “Subject + Location + Event + Genre/Style + Camera movement…”
3. Detailed Plot Description. A shot sequence or a timeline, either is accepted, with each segment’s visuals, camera movement, action, dialogue and sound effects described.
4. Additional Notes. Picture details that run throughout: camera position, environment, sound, atmosphere.

Numbering your references

Seedance 2.5 takes up to fifty reference assets in one request, and the guide’s position is that the binding between an asset and its role in the story is where multi-reference work succeeds or fails.

As the number of reference assets increases, the mapping and reference relationships between assets become especially important. It is not recommended to provide mapping information only inside the image itself.

The failure it names: write “John” on a character’s reference image, then write “John is at school” in the prompt, and you get duplicated or confused characters. Numbering follows upload order, and every asset gets bound in the text.

Three notations appear across the guide’s own examples, and all three are official:

The Chinese edition also shows a compressed form for grouping, img1-2 是人物 1，对应音频 1: images one and two are character one, who uses audio one. The English edition renders this as “Images 1-2”.

Two binding patterns are worth copying directly. For a character with a voice: “Image 1 is the protagonist Zhang San, and Image 1 uses the timbre of Audio 1.” For borrowing motion from footage: “Strictly refer to the action and camera movement in Video 1, keeping the order consistent with the video.”

Shots, timestamps and the continuity rule

Timestamp response is the headline change from Seedance 2.0. The guide states it directly: Seedance 2.0 “does not respond to timestamps and only responds to shot numbers, while Seedance 2.5 supports integer-second timestamps.”

Thirty seconds, divided

Both dividers are accepted, and they can be combined. Shot 1: headers work, integer-second ranges like 0-3s: work, and the guide’s own worked examples use forms including Shot 3 (6-10s):. Three time-control modes are supported: clear intervals, specific time points, and relative time.

The sizing advice is practical. Under-fill a range and the model invents to cover it; over-fill one and you get excessive cutting or dropped plot.

Locked and unlocked tasks

This concept did not exist in Seedance 2.0, and it explains a class of confusing failure. Every request is sorted into one of five task types, and the type decides whether you still control the output’s shape.

Whether your input takes control of the output

For first and last frames there are two routes with different consequences, and the guide is explicit about the trade. Setting the image role through the first_frame or last_frame parameter locks the output’s aspect ratio to the supplied image. Setting the role to reference_image and naming the frames in the prompt text (“Image 1 is the first frame.”) does not lock the ratio, and the result will be similar to the reference rather than an exact match.

Camera language it already knows

The guide names its accepted vocabulary rather than leaving you to guess, which is the single most useful page in it. These terms can be written directly:

For anything outside that list, the instruction is to convert it into [term + descriptive explanation]. The guide’s own example: “Rack focus: the focus of the frame shifts smoothly, the trees in the foreground that were sharp become blurred, and the figure in the background gradually comes into focus.”

Transitions get the same treatment, with the trigger point and the method both written out. The Chinese edition’s example is 第 5s 快速向左横移转场（向左擦除+自然叠化）: at five seconds, a fast leftward lateral transition, wipe left plus a natural dissolve.

Audio, dialogue and subtitles

Two notations exist across the pages published that morning, and they are not the same. The prompt guide writes dialogue inside a shot block:

Shot 3 (6-10s): Facial close-up. The elderly woman looks reluctant to part.
Dialogue (elderly woman): "Fly safe, my child. Come back to me."

The companion tutorial gives a bracket scheme instead: () for music, <> for sound effects, {} for dialogue and 【】 for subtitles, with a recommendation to name the language before any non-Chinese dialogue. Both come from ByteDance, published within a minute of each other, and neither page mentions the other. The shot-block form is the one the guide’s worked examples actually use.

Audio is on by default. The generate_audio boolean defaults to true, and with it the model produces matching voice, sound effects and background music from the prompt and the visuals together. Lip sync is demonstrated in ByteDance’s own multilingual example rather than stated as a specification.

Negative control is deliberately narrow

There is no negative-prompt field. Negation is written into the prompt, and the guide sanctions it for two things only: subtitles and audio.

The API is 720p, whatever the marketing says

Seedance 2.5 does not generate 4K through the API, and it does not generate 1080p. BytePlus’s reference is unambiguous: “Dreamina Seedance 2.5: Default 720p; supports 480p and 720p.” The resolution table tags 1080p as unsupported on 2.5 and restricts 4K to Seedance 2.0 alone. On maximum resolution the API for 2.5 sits below the API for 2.0.

The 4K and 60 fps figures in circulation come from the consumer Dreamina product page, which sells both as premium subscription features alongside a beta mode that stretches a clip to 180 seconds. None of the three appears in the model API. ByteDance has not published how the consumer 4K path relates to the model’s own output, so read a 4K claim as a statement about the Dreamina product rather than about Seedance 2.5 itself.

The limits, as documented

At 720p the pixel dimensions are 1280×720 for 16:9, 720×1280 for 9:16, 960×960 for 1:1, 1112×834 for 4:3, 834×1112 for 3:4 and 1470×630 for 21:9.

Two restrictions are easy to trip over. Reference images and videos containing real human faces cannot be uploaded to Seedance 2.5 or the 2.0 series. And multi-panel storyboards work best at fifteen panels or fewer: an eighteen-panel input is given as the case that produces still frames or scrambled sequence order.

Start with the skill, not the prompt

Before any prompting advice, the guide’s first section tells you to install a prompt-optimisation skill that ByteDance hosts on its own CDN.

Model names and where it runs

One model, no pro, lite or fast tiers. The name differs by platform, and the version stamp is the same on both:

Rate limits on BytePlus are 600 requests per minute with 10 concurrent tasks for enterprise accounts, and 180 requests per minute with 3 concurrent for individuals. The offline flex inference tier is not supported for this model.

Volcano Engine is effectively closed to a UK sole trader: activation needs Chinese real-name verification and a prepaid renminbi balance. BytePlus ModelArk is the international door, and its public endpoint runs in Asia Pacific, so anyone with a data-residency requirement should settle that before building on it.

The resellers, priced

Seedance 2.5 shipped consumer-first on Jimeng AI and Doubao Pro on 31 July, with the API following. Western aggregators picked it up quickly, and their prices for identical output are a long way apart. Every figure below is from the provider’s own pricing page, checked on 7 August 2026.

Segmind runs at roughly half fal.ai’s rate for the same 720p output. At Segmind’s price a 30-second clip, the model’s maximum, costs about $7.17; the same clip on fal.ai is about $14.19. Smaller resellers advertise lower headline rates still, so check the current page before committing to a route.

The OpenRouter row is the one to read twice if you buy inference through credits there. OpenRouter sells video generation and it sells Seedance, but the newest version it carries is 2.0. Seedance 2.5 needs an account somewhere else.

Three things to settle before you build on it

The consumer route and the developer route have different licences. BytePlus’s specific terms for video generation treat output as your data and claim no ownership of it, while barring resale of the generation service itself. Dreamina’s consumer terms give you nominal ownership but take a perpetual, irrevocable, sub-licensable licence that extends to other users of the platform, and they disclaim any uniqueness in what you generate. They are also scoped to United States users and name no UK contracting entity. For brand assets, that is the difference between a usable licence and an unusable one.

Provenance is documented for 2.0 and undocumented for 2.5. ByteDance said in March 2026 that Seedance 2.0 output carries C2PA content credentials and an invisible watermark. Its own Seedance 2.5 page says nothing about watermarking, content credentials or labelling, and no 2.5-specific statement has been published. China’s AI content labelling measures, in force since 1 September 2025 under standard GB 45438-2025, bind the Chinese platforms, so Jimeng and Doubao output carries a visible label that the international route may not. Test the actual file rather than assuming.

Inference runs in Asia Pacific. BytePlus ModelArk’s public endpoint is ark.ap-southeast.bytepluses.com, and BytePlus moved that region from Singapore to Johor while keeping the old display label, so the region name does not tell you where compute sits. Its data authorisation agreement allows storage outside the region of use. There is no UK or EU inference region.

What ByteDance still has not put a number on

Three things are genuinely open, and both of the first two are contradicted inside ByteDance’s own materials.

How far a video can be extended. The Seed project page says a clip can be extended “twice”, implying a ceiling near 90 seconds. The API documentation describes extension “over multiple rounds” with no cap. The launch blog claims output “lasting several minutes”. No authoritative figure has been published.

How many shots fit in thirty seconds. Shot count is not a parameter. It is controlled through prompt text alone, and no maximum is stated.

How Seedance 2.5 is built. Seedance 1.0 and 2.0 both have arXiv papers. Seedance 2.5 has none as of today, so architecture, parameter count and training data are undocumented. The launch blog says only that it builds on 2.0’s unified multimodal audio-video joint-generation architecture.

Read the Chinese edition if you can

The two editions are not identical. The Chinese version links a further ByteDance document, Seedance 2.5 提示词模板 (Seedance 2.5 Prompt Templates), hosted on Lark, which the English edition omits entirely. That link currently redirects to a login page and is not publicly readable, so its contents cannot be verified, but its existence tells you there is a per-task-type template set the English guide does not point at.

The English edition is longer in raw text, at roughly 64,000 characters against 29,000, because it inlines fuller example prompts. For the examples, read English. For the complete set of pointers, read Chinese.

---

For the model itself, its lineage and how it ranks against rivals, see Seedance 2.0 and 2.5. The wider field is in our AI video models hub, and the toolchain for driving these models from a written brief is in animate and edit video with Claude and MCP. New terms are defined in the glossary.