seedance-25

v2.0.0 · 2026-08-09 · MIT · see ATTRIBUTION.md

Prompt engineering for Seedance 2.5 on any surface. The deliverable is always a copy-paste prompt plus the settings block the operator sets, whether that surface is a web UI, a desktop app, or an agent. This skill writes prompts. It does not generate, upload, or send anything.

What this skill is

Seedance 2.5 doubled the clip length and more than tripled the reference budget. Both changes reward people who write like a director and punish people who write like they are filling in a search box. More room does not mean you can be vaguer. It means there are more places to go wrong before you find out, and you only find out after the render finishes.

This skill is the discipline that keeps that from happening.

This skill is self-contained. It does not depend on any other skill being installed.

Precedence when files disagree: official-spec.md → surface-profiles.md → everything else. The first stands on ByteDance documentation published 2026-08-07; several other files were originally written from practitioner reports. Where they conflict, the documentation wins, and the other file should already say so. If you find a contradiction that has not been reconciled, say so out loud rather than picking one.

Evidence grading, and why it exists

Seedance 2.5 was announced 2026-07-31 and documented 2026-08-07. A lot of what is written about it online is still launch-window reporting or vendor marketing. Every factual claim in this skill carries a label:

Never state an unlabeled number as fact. If a value is not confirmed for the surface in play, write it into the settings block with (verify in UI) next to it. Guessing a resolution or a reference ceiling is how a client gets promised something the tool cannot do.

[community] is usable as a working method. It is not usable as a claim in a client deck or a sponsored post.

Step 0: identify the surface. Always.

Without a surface there is no settings block and half the limits are invented. 2.5 runs on several products. Each has a different UI, different tag syntax, different resolutions, and different feature availability. Before writing a single sentence:

1. Ask where the prompt is going. Do not default.
2. Load that surface's profile from references/surface-profiles.md.
3. If the surface is not in the profiles, use the conservative profile (also in that file), mark unverified values as unverified, and say out loud that the profile is conservative. Never carry a limit from one surface to another, not even between two ByteDance products.

Hard dividing line: 30 seconds in one generation and 50 references is Seedance 2.5. Seedance 2.0 is still 4-15 seconds and 9 images / 3 videos / 3 audio. Never write "Seedance 2.0 now does 30 seconds." It is wrong and it is the single most common mistake made about this model.

Second hard line, and it is new: [official 2026-08-09] Seedance 2.5 runs at 480p and 720p only. It does not do 1080p and it does not do 4K. Anywhere. Not on Magnific, not on Higgsfield, not in CapCut, not in Dreamina, not through the API. This is a property of the model, not a limitation of a surface. Only Seedance 2.0 reaches 4K, and it does so at 15 seconds. Earlier versions of this skill said the opposite and told people to go to CapCut for higher resolution. There is nothing there either.

The model ID is public: dreamina-seedance-2-5-260628 (480p/720p, 4-30s, 24 fps, mp4 and mov). The API is not enterprise-only; the individual tier allows 180 RPM and a concurrency of 3. Full model card, rate limits, and pricing are in references/surface-profiles.md.

Thirty seconds is the documented ceiling for one generation. Video Extension is real and documented, but the official docs give no ceiling for the final cut, and the 60s figure is community-reported. Ultra-Long mode does not appear in the documentation at all, and the model card says 4-30s. When someone asks for longer than thirty seconds, load references/modes-and-workflows.md and offer Extension, and when quoting any number above 30 seconds, say that it is a community claim the official model card contradicts.

The trade is length against resolution. A client master at 1080p or 4K is Seedance 2.0 at 15 seconds, with no exceptions. Length, a large reference budget, timestamps, or editing finished footage is 2.5 at 720p. Say which one you are on before you hand over the prompt, not after the render.

Model invariants

Do not take these from a UI. They are properties of the model:

• Prompt spine: subject → action → camera → light and setting → style → sound. The official macro-structure wraps it in four blocks (asset referencing → one-sentence summary → detailed plot → additional notes), see official-spec.md section 4.
• At 30 seconds only one slot changes: the action holds a short arc, not a single motion.
• Write in stages, not one paragraph. One primary change per stage and an explicit End state: for each. That end state is the anchor the next stage attaches to.
• Timestamps work, but only on 2.5. Integer seconds, no gaps in the timeline, never for high-frequency actions. Seedance 2.0 ignores them entirely and understands only shot numbers. [official]
• Locked tasks set their own aspect ratio and duration. Editing, first/last frame, and extension inherit those from the input asset, so there is nothing to ask the operator about. [official]
• Time windows are budgets, not frame-exact cuts. Underfeeding a beat does not trim it, it rushes it.
• Pin the constants once at the top and restate them once at the bottom as a consistency line.
• The audio budget does not scale with duration. The limit is per line and per breath, not per clip. A 30-second clip is room for more lines, not permission for longer ones.
• An unbound reference blurs the result. One dimension, one owner, exclusions written out.
• Positive description is preferred. Negatives are officially supported only for subtitles and audio, not for visuals. [official]
• Emotion is described through visible signals, never named. Never "she is sad."
• Anti-slop applies everywhere. A specific lens, a specific movement, a specific light source.

Always surface-specific and resolved from the profile: tag syntax, reference ceiling, available resolutions, aspect ratios, whether duration is continuous or a fixed set, whether audio references exist, whether Extension and region-level editing are present, and what any of it is called in the UI.

Beat architecture for 30 seconds

ByteDance's own guidance splits 30 seconds into four beats. Treat it as the default skeleton, not dogma. It holds on every surface because it is a property of the model, not the UI:

The rules that hold it together:

1. One arc, not four ideas. Beats are phases of a single thought. If every beat introduces a new subject, the model glues them into a trailer collage.
2. The callback in beat four is not decoration. Returning to an object or gesture from beat 1 is what gives a 30-second clip closure instead of an ending that just stops.
3. Roughly 4-6 seconds per beat is comfort, not a ceiling. Four beats sit well in 30 seconds. So do eight and nine. The official 30-second examples in ByteDance's own documentation run 8 and 9 shots at 2-4 seconds each [official 2026-08-07]. The earlier "seven is the ceiling, eight compresses" line in this skill was a guess and is withdrawn. What breaks a beat is too much plot inside its window, not the number of windows. When in doubt, add a window and take something out of each, rather than forcing two changes into one.
4. Pin the constants once, at the top. Subject description, wardrobe, props, lighting mode, and grade get stated once and declared to hold for the whole clip. Repeating them in every beat is filler, but never stating them at all is the most common cause of drift at 30 seconds.
5. The end of a beat is a completed action. A beat's sentence lands on impact and the next beat opens something new. Otherwise the cut falls in the middle of a movement.
6. Every beat gets an End state:. The beat describes what happens, the end state says where it should land. Without it, stages blur and the model invents the join. This is the cheapest edit with the largest effect on a 30-second clip.

Both at once, the beat as dramaturgy and the end state as a machine anchor:

[0-8s] A florist trims stems behind a workbench.
End state: she holds the finished bouquet in her left hand.

[8-16s] She wraps it in kraft paper and ties a green ribbon.
End state: the wrapped bouquet lies centered on the workbench.

Keep her identity, clothing, and the workbench layout consistent throughout.
No subtitles, no background music.

Skeletons and worked examples: references/beat-templates.md. Longer formats, extension, and editing finished footage: references/modes-and-workflows.md.

Reference binding

An unbound reference is a reference that blurs. Every asset gets one primary role and an explicit exclusion. Tag syntax is surface-specific (@Image1 in most UIs, a references[] array in an API), the rules are the same everywhere:

@Image1 defines the main character's identity: face, hair, wardrobe. Do not take background or composition.
@Image2 defines the location and colour palette. Do not take any person.
@Video1 defines camera motion and pacing only. Do not take appearance, wardrobe, or location.
@Audio1 is the music bed. Beat 3 lands on the drop.

Hard rules:

• One dimension, one owner. Identity, motion, camera, environment, timing, style. If two assets control the same dimension, the model averages them and you get neither.
• Exclusions are written, not assumed. "Motion only, no appearance" is a functional part of the prompt, not a comment.
• 50 slots is a trap as much as a feature. Thirty unlabeled images are thirty ways to average your subject into mush. More slots raise the ceiling on role separation, not on dumping.
• The quality band sits far below the ceiling. [official] 1-8 distinct subjects from images (stretch 9-12), 1-5 from video (stretch 6-10), donor clips 5-10 seconds, edit sources ≤20 seconds, 1-5 reference images for an edit. Past those numbers it turns into a slot machine. Full table in references/official-spec.md section 8.
• Multi-view inputs are allowed on 2.5 [official] and not on 2.0. Up to 5 subjects, multi-view is fine. Above 5, prefer one view per subject and split extra angles into separate images. One image holding several viewpoints is wrong in every case; the model reads it as one composition.
• A tag's number is its upload order [official], not the order it appears in the prompt.
• Mapping does not go inside the picture. [official] Writing a character's name onto their reference image and then using that name in the prompt is a documented route to character confusion and duplication. Bind in the text.
• When a reference is accurate, do not describe the scene again. [official] "Strictly refer to the actions and camera movements in Video 1" is enough; spelling out its contents makes it worse.
• Tags are never translated or renumbered. Preserve exactly the shape the surface or the operator supplies. But there is no single shape: the official documentation uses @Image 1 with a space, [Video 1], <video1>, and Images 1-2. The earlier "no spaces" rule in this skill was invented and is withdrawn.
• An asset that owns nothing gets dropped.
• The ceiling is surface-specific. The model does 50: 30 images (up to 4K each), 10 videos ≤30s combined, and 10 audio clips ≤30s combined [official]. A given UI may allow fewer.

Group ranges are official [official]: Use Images 1 to 7 in order as keyframes. and Images 1-2 are Character 1 and correspond to Audio 1; Images 3-4 are Character 2 and correspond to Audio 2. A group still needs one role and one exclusion, stated once for the group.

The tag @ClayRender1 does not exist and never did. It was invented in an earlier version of this skill. 3D clay-model reference is a real, documented task, but it binds with an ordinary [Video 1] plus a sentence naming what is taken from it: "Refer to the camera movement and motion in [Video 1]." See references/official-spec.md section 10.

Output contract

Deliver the same shape every time so the operator can paste and go:

1. The final prompt, in English.
2. A settings block, because the prompt alone does not determine the clip. Fields come from the surface profile and the heading carries the surface name:

   --- SETTINGS: {surface} ---
   Mode: {from profile}
   Model: Seedance 2.5
   Duration: 30s        (locked tasks: inherited from the source)
   Aspect: 16:9         (locked tasks: adaptive, inherited from the source)
   Resolution: 720p     (2.5 does not go higher)
   Format: mp4          (editing and extension: mov)
   References: @Image 1 = ..., @Video 1 = ...

   Any value the profile has not confirmed goes in with (verify in UI). Do not invent it. On the API, add content.role per asset plus ratio and duration according to the task type.
3. State the resolution before anyone asks. 2.5 is 480p or 720p, full stop. If the prompt is aimed at a 1080p or 4K master, say so and let the operator choose: 2.5 for length, references, and timestamps at 720p, or 2.0 for 15 seconds at 4K. Never hand over a 720p render as if it were what was ordered.
4. Do not silently run it. Magnific and Higgsfield can generate 2.5 through an agent [verified-live 2026-08-07], but generation costs credits and is the operator's decision. On Higgsfield, 2.5 also has no keyframes, so anything built on a start/end frame stays on 2.0 or goes manual.

Procedure

1. Identify the surface. Step 0. Load the profile, resolve limits and syntax.

2. Pick the task and establish whether it is locked. [official] This comes before asking about aspect ratio and duration, because on a locked task there is nothing to ask:

   Situation	Task	Locked?
   no assets	text to video	no
   one starting image	image to video / first frame	yes, ratio from the frame
   locked opening and closing image	First & Last Frame	yes, ratio from the first frame
   several assets with roles	Omni Reference / reference-to-video	no
   output must match a board shot for shot	Keyframe reference	no
   loose guide, rough plot only	Storyboard reference (≤15 panels, line art)	no
   a coarse 3D blockout carries motion and camera	3D clay-model reference	no
   changing something inside finished footage	Editing (Smart Edit)	yes, ratio and duration
   extending existing material	Extension	yes, ratio from the source
   joining two finished clips	Seamless transition	from the sources

   Keyframes versus storyboard is the choice people get wrong most often. A storyboard is loose and the output will not follow it panel for panel. Keyframes are strict. When a client wants their board reproduced exactly, that is keyframes. Full catalogue: references/official-spec.md section 7.

3. Establish intent and format. What the clip is for and where it is going. On unlocked tasks, always ask for aspect ratio and duration and never default them; the ratio can be anything between 0.4 and 2.5 [official], so you are not stuck with 2.0's six fixed values. On locked tasks, do not ask. Just state what is inherited from the input.
4. On locked tasks, set the parameters and put a trigger word in the prompt. [official] ratio = adaptive on all three, plus duration = -1 for editing, plus output_format = mov for editing and extension. Editing needs add / remove / replace / change to / modify in the prompt; extension needs continue / extend forward / extend backward. Without that word the model cannot tell which task you mean. With more than one input video, name the specific one or the model chooses for you. For extension and editing, load references/modes-and-workflows.md.
5. Safety gate. A real face, a third-party brand, a public figure, a protected character: resolve it through the IP gate in references/craft-essentials.md before writing the first sentence.
6. Build the beats. One arc, callback at the end, End state: on each. Four windows is a comfortable default; eight or nine is equally fine. On 2.5, write integer-second timestamps with no gaps in the timeline.
7. Pin the constants. Subject, wardrobe, light, grade, ambience. Once at the top, restated once at the bottom.
8. Bind the references. A role and an exclusion for every tag, numbered by upload order. If the references do not exist yet, build them before writing the prompt (modes-and-workflows.md section 6): lock one look, generate by category, build reference sheets for anything recurring, then leave them alone.
9. Run anti-slop. references/craft-essentials.md. No "cinematic", "stunning", "breathtaking", "ethereal", "glow", "majestic".
10. Deliver per the output contract. Prompt, then settings block with the surface named.
11. Say draft-first out loud. At 30 seconds a bad take costs double. Test the look on a short clip, then commit the length.

Pre-flight checklist

• Surface identified, profile loaded, unconfirmed values marked (verify in UI)
• Resolution matches the model: 2.5 is 480p or 720p. If the client wants 1080p or 4K, that is 2.0 at 15s
• Established whether the task is locked or unlocked
• Unlocked: aspect ratio and duration confirmed by the operator, not guessed (ratio 0.4 to 2.5)
• Locked: ratio = adaptive, plus duration = -1 for editing, plus output_format = mov for editing and extension
• Editing or extension carries a trigger word, and the specific video is named when there is more than one
• Mode matches the assets and is named the way that surface names it
• Strictness matches the brief: keyframes when the output must match exactly, storyboard when a loose guide will do
• Reference ceiling matches the surface, not the model maximum
• Videos total ≤30s and audio totals ≤30s, not merely each clip under the limit
• Every tag has a role and an exclusion, numbered by upload order
• Mapping lives in the text, not written inside the images
• No dimension has two owners
• References inside the quality band (1-8 image subjects, 1-5 from video, clips 5-10s), not merely under the ceiling
• Subject and main action stated plainly, without ornament
• Constants stated once at the top and restated at the bottom
• Every stage has one primary change and an explicit End state:
• The timeline is continuous, with no gaps between windows
• Character count, wardrobe, and prop ownership consistent across stages
• The final beat contains a callback
• Lines are short, one thought per breath
• Emotion written as visible signals, not named
• Actions described generally, only a few detailed, none repeated
• Negatives limited to subtitles and audio; visuals described positively
• Forbidden items listed at the end (subtitles, background music, morphing, extra characters)
• If the track must be clean: a [SOUND] directive, not just "no music"
• On-screen text carries a language directive wherever text appears
• Prompt passed anti-slop
• IP and likenesses clean or authorized
• Settings block attached

The exact wording of the repair lines referenced above lives in references/proven-fixes.md. Paste them as written. A reworded line is a different prompt.

Two notes on talking about this model publicly

Do not oversell it on the vendor's behalf. ByteDance's own documentation states that 2.5 is not a generational leap over 2.0 in the way 2.0 was over 1.5, and describes it as a systematic hardening for production work. When the vendor is that measured, inflating it in a post is a risk with no upside and it is trivially checkable. If you are writing sponsored or ambassador content, that applies twice, alongside the usual disclosure and verifiable-claims discipline.

ByteDance publishes its own prompting skill for 2.5 (/sd25-pe, installed via npx skills, see references/official-spec.md section 12). This skill is not the only option and should not be presented as one. Installing theirs is the operator's call.