Seedance 2.5 Prompts: a Practical Guide

A good Seedance 2.5 prompt names five things, anchors the ones that must stay stable, and leaves the rest open. A bad one either describes nothing specific or tries to specify everything, including controls that do not exist. This guide covers the structure that works, the vocabulary that lands, and the requests that waste credits because the platform does not take them.

Product behaviour described here follows the product team's confirmation dated 14 October 2026.

The five-part structure

Every reliable prompt covers subject, action, camera, setting and mood.

For the bigger picture, read our overview of ByteDance Seedance 2.5 AI video generation.

A courier in a red jacket pauses at a rain-soaked station, lifts an envelope and looks toward an arriving train; slow camera push-in, blue evening light, quiet anticipation.

Subject: the courier, and her jacket. Action: pausing, lifting, looking. Camera: a slow push-in. Setting: a wet station at dusk. Mood: anticipation, quiet.

Drop any one of those and the model fills the gap with something plausible but arbitrary, and it fills it differently on each attempt. That inconsistency between attempts is usually read as the model being unreliable, when it is the prompt leaving a decision open.

Anchor what must not change

After the five parts, name the details that have to stay stable: the jacket's colour, the product's silhouette, the direction the light comes from, the logo's placement.

If you are comparing platforms, here is where you can use Seedance 2.5 today.

This is the single highest-leverage habit. The model holds what you name and improvises what you leave open. If character identity matters across several shots, go further than words: generate the character as its own asset first, then reference that asset in every shot. Re-describing a person in each prompt is how faces drift.

Describe the things that are not settings

Seedance 2.5 has no numeric controls for outline weight, shadow intensity or motion speed, no frame rate or frame sampling, no keyframes, no frame interpolation and no motion presets. None of that is a gap in your prompt; those controls simply are not part of the product.

New to the model? This walkthrough covers how to access Seedance 2.5 step by step.

So those qualities go into the language instead:

Written this way, the request is something the model can act on. Written as "set motion speed to 0.7", it is a request for a control that does not exist, and the attempt is wasted.

Timing: use timestamps, not a beat grid

Precise timestamped instructions are supported, and they are the right tool when a specific beat must land at a specific moment. Keep every timestamp inside the duration you selected for the generation.

You can try the ByteDance Seedance 2.5 video generation model directly on OiiOii.

What is not available is a tempo map, beat markers, or a manual time grid you edit afterwards. The distinction is simple: you say when something happens in the prompt; you do not adjust it on a timeline later. For lyric work, subtitle appearance timing is separately editable, which is a different thing again and worth not confusing with beat control.

A prompt should fit what the model can render in one shot, and our Seedance 2.5 model guide lists those limits, from shot duration to reference inputs.

Match the prompt to the duration

Each generation runs 4 to 30 seconds, and you choose. The prompt should describe an arc that fits.

Our step-by-step guide explains how to make a music video with AI tools.

A single gesture written for 30 seconds produces a shot that idles, because the model has more time than action. A three-stage sequence compressed into 5 seconds produces a shot where none of the stages read. Pick the duration from the action, then write to it.

For anything past 30 seconds, you are not writing one prompt. Story anime generates multiple storyboard shots and assembles them automatically, so a 90-second piece is a shot list, each with its own prompt, reviewed at the storyboard stage before generation.

Using references well

Up to 50 reference inputs are allowed in total, combining images, video and audio. The limit is combined, not 50 of each type.

Quantity is not the lever. A small, internally consistent set beats a large contradictory one every time, because contradictions force the model to choose and it will not choose consistently. Tell it what each reference is for: this is the face, this is the room, this is the palette, this is the atmosphere of the sound.

Audio is generated by default and can be turned off. If you are supplying an audio reference, say what it is meant to establish, mood, pace, instrumentation, rather than assuming the model will infer the role you had in mind.

Prompt patterns that work

The stable-subject pattern. Lead with the subject and its fixed traits, then the action, then everything else. Useful when identity consistency is the priority.

A tall woman with short silver hair and a grey wool coat, collar up, walks between two market stalls and stops to look at a cage of paper lanterns; camera tracking beside her, overcast afternoon light, curious and a little lost.

The single-beat pattern. One action, one camera move, short duration. The fastest way to learn what the model does with your subject matter, and the cheapest to iterate on.

A copper pan lifts off a gas flame and tilts toward the camera; static camera, warm tungsten light, late-service calm.

The environment-first pattern. When the place is the point and the subject is incidental.

An empty rooftop pool at dawn, steam rising off the water, a single towel left on a chair; slow drift to the left, pale blue light before sunrise, abandoned stillness.

The reference-anchored pattern. Minimal description, heavy anchoring, used when you have already generated the character and scene as assets.

The character from the attached reference enters the attached room, sets a box on the table and steps back; handheld camera, matching the reference lighting, apprehensive.

Revision prompts

Changing the whole prompt between attempts means you learn nothing about which element mattered. One change per generation, and keep a short log of model, duration, references, the change, and what it did.

When something specific is wrong and the rest is right, local region editing lets you correct that region without regenerating the shot. Green screen and white model editing are also supported. Reaching for these before a full regeneration is the difference between a project that finishes on budget and one that does not.

Within the story workflow, you can also address a named agent for a revision. The seven are Art Director, Scriptwriter, Scene Designer, IP Designer, Character Designer, Storyboard Artist and Sound Director, the system assigns work between them by default, and the roster is fixed. When the current set does not seem to cover what you want, the move is a more specific creative request rather than a new agent.

What a prompt cannot ask for

Writing these into a prompt produces a wasted generation, so it is worth knowing the list:

• Resolution above 1080p, or a 10-bit toggle. Output is 1080p and 10-bit is built in.
• Frame rate, frame sampling, keyframes, audio channels, interpolation or motion presets.
• Beat markers, tempo maps or a manual time grid.
• Manual mouth shapes, phonemes or lip-sync keyframes. Lip sync is automatic and unrestricted by mode or model; if the result is off, review it and ask for a revision.
• Separate background, character and motion layers to composite afterwards, or an animatic or previs pass.
• Transition effects, colour grading or added effects. The editor handles shot order, trimming and compositing.
• Manually added score, voiceover or sound effects, or volume and mixing adjustments. Audio is generated.
• A multi-format export. Aspect ratio is chosen before generation, and a second format means a second generation.

Style filters and shaders, on the other hand, are supported, so style-level direction is a legitimate request.

A practice routine

Take one subject you care about. Write the five-part prompt. Generate at a duration that fits the action. Watch the whole result.

Then run four more generations, changing exactly one thing each time: the camera move, the mood word, the duration, the reference set. Keep all five prompts and all five results side by side. That hour will teach you more about how the model reads your material than any general guide, including this one.

Before and after: four prompts rewritten

Nothing teaches prompt structure like watching a weak one get repaired.

Before. "A cool cyberpunk city scene, very cinematic, high quality, 4K."

Three problems. There is no subject and no action, so the model invents both. "Cinematic" and "high quality" describe nothing it can act on. And 4K is not available, so that word is doing no work at all.

After. "A courier on a parked motorbike checks a handheld screen while rain runs off a neon sign above her; camera slowly pushing in from street level, magenta and cold blue light pooling on wet tarmac, wary stillness."

Before. "Make the character move faster, set motion speed higher."

The problem is a request for a control that does not exist. There is no motion speed parameter.

After. "She crosses the room in three decisive strides rather than strolling; the movement snaps forward and stops hard."

Before. "Same character as last time, in a kitchen."

"Same character as last time" is not something the model can honour across generations. It has no memory of your previous shot.

After. Attach the generated character asset as a reference, then: "The character from the attached reference leans against the kitchen counter and reaches for a mug just out of easy range; camera static at counter height, flat morning light from the window behind her, half-awake."

Before. "Cut to the beat at bars four and eight."

There are no beat markers, bar-based cut tools, tempo maps or manual time grids, so this cannot be acted on.

After. Use timestamped instructions inside the shot's duration: "At around 4 seconds she looks up sharply; at around 8 seconds she turns away from the window." Timestamp prompting is supported and precise.

A prompt template you can reuse

Keep a file with these five lines and fill them in each time. The discipline matters more than the wording.

SUBJECT:  who or what, plus the traits that must not change
ACTION:   one clear thing that happens, with a beginning and an end
CAMERA:   one move, or deliberately static
SETTING:  where, when, what the light is doing
MOOD:     the feeling, in two or three words
ANCHORS:  the specific details that must survive (colour, shape, placement)

Then, separately, record the mechanical choices so your log is useful later:

DURATION:    4-30s, chosen from the action
RATIO:       chosen from the destination, unchangeable afterwards
REFERENCES:  what each one establishes
TIMESTAMPS:  if any beat must land at a particular moment

Articles and guides come and go; a filled-in version of this template for twenty of your own shots is a permanent asset.

Frequently asked questions

How long should a prompt be?
Long enough to cover the five parts and name the anchors, and no longer. Extra adjectives that do not constrain anything add noise rather than control.

Should I write in sentences or keywords?
Sentences. The model reads relationships between subject, action and camera, and keyword soup discards those relationships.

Can I reuse one prompt across several shots?
Reuse the visual-language lines verbatim, so the look stays consistent, and change the subject-action-camera part. That combination is what produces a coherent sequence.

How do I keep a character consistent?
Generate the character as an asset and reference it, rather than describing them in each prompt. This is more reliable than any phrasing.

Why do two identical prompts produce different results?
Anything you did not name is a decision the model makes freshly each time. Naming the anchors narrows that space.

Can I specify a frame rate or an animation curve?
No. There are no frame rate, frame sampling, keyframe, interpolation or motion preset controls. Movement quality is described.

Can I ask for 4K or set 10-bit?
No. Output is 1080p currently, and 10-bit is built in rather than switchable.

Can I ask for a specific transition into the next shot?
No. Transitions are not an editor function. What reads as a transition comes from how adjacent shots were described and ordered.

Can I ask for specific sound effects at specific moments?
Audio is generated. You can describe the intended sound as part of the scene, but score, voiceover and sound effects cannot be manually added or mixed.

Can I prompt for a style?
Yes, and style filters and shaders are supported as well, so style-level direction is available on top of the language.

Next steps and official resources

For the end-to-end working method, see how to use Seedance 2.5. For a full production walked through, see the Seedance 2.5 review. For what each generation costs, see the Seedance 2.5 pricing guide and the pricing page. For the platform context, see the OiiOii workflow overview.

Product details in this guide follow the product team's confirmation dated 14 October 2026; promotional rewards can change.

Related guides

• ByteDance Seedance 2.5 AI video generation
• ByteDance Seedance 2.5 video generation model
• Where can I use Seedance 2.5
• How to access Seedance 2.5
• Seedance 2.5 pricing