Seedance Prompt Builder

Seedance gives you a lot to work with: multi-shot clips, native audio and a generous reference system. That is also what makes a vague prompt expensive, because the model has more freedom to wander. The builder above turns a brief into a prompt that pins down the subject, action, setting, camera and sound, and the notes below explain what is particular to Seedance and what changed in 2.5.

It runs in your browser. Nothing is uploaded, there is no account, and it costs nothing, because it writes prompts rather than video.

Which Seedance you are writing for

Seedance is ByteDance’s video model family, and two versions are in use at once.

Seedance 2.5 is the current release. ByteDance announced it on July 31, 2026: 30-second clips in a single pass, multiple rounds of extension, up to 30 images, 10 video clips and 10 audio clips as references, and timestamp-level control for editing audio and video. ByteDance says the joint audio-video design also cuts down on subtitles and background music appearing unasked. The same day, PR Newswire reported that Dreamina launched it worldwide, with global availability beginning on Dreamina. The CapCut newsroom says CapCut’s rollout starts the following week for paid subscribers over 16 in Europe, Asia, the Middle East and South America, and it does not name the US in that first wave. Runway and PixVerse also run it: Runway on its web app, PixVerse for Pro, Premium and Ultra members with Video, Transition and Reference modes. In China it runs in the Jimeng AI and Doubao apps.

Seedance 2.0 is still the broadest option. ByteDance’s Seed team launched it on February 12, 2026. It reached CapCut from March 26, 2026, starting with Brazil, Indonesia, Malaysia, Mexico, the Philippines, Thailand and Vietnam, then Africa, South America, the Middle East, Europe and Japan, and the US from April 7, according to the CapCut newsroom. If you are a CapCut user, especially in the US, 2.0 is the version in front of you today. It is also on Dreamina, PixVerse, Pollo AI and Runway.

This builder lets you pick either. The sections below mark which advice is for which version.

How prompting Seedance differs

Length and shape depend on the version. For 2.0, Dreamina’s guide recommends 30 to 100 words and explains why: too brief and the output is unclear, too long and the focus drifts. Its guide for the lighter Seedance 2.0 Mini says roughly 25 to 45 words. For 2.5 the clip can run to 30 seconds, and Runway’s 2.5 prompt guide lays a longer prompt out in four parts: an opening list that numbers each uploaded asset and binds it to a role; a one-sentence summary of subject, location, event, style and camera movement; a timeline of gap-free segments with integer-second timestamps; and a closing line of consistency notes for details that must not change. Runway adds that its own output treats the timestamps as pacing guidance, not frame-exact cuts, and suggests roughly 6 to 8 seconds per beat as a sensible default. That is Runway’s guidance for its platform, not a ByteDance specification; it is the best published structure for 2.5 that we found.

Order: subject first. Dreamina’s advice is to lead with the main subject or character, add the visual details such as style, lighting and composition, then the action. Its Mini guide lists subject with style, a specific action, camera language, setting with lighting, then mood and pacing. The builder follows the same spirit: who, what they do, where, how it is filmed, and what you hear. Dreamina also says to focus on what affects the video and leave out details that change nothing.

Camera: name one move, and know what changed. ByteDance describes Seedance 2.0 as doing prompt-driven camera planning, meaning it plans camera language for you when you do not specify it. If a particular move matters, say it once and clearly. Runway’s table of differences says 2.0 takes one camera movement per shot, while 2.5 allows compound moves and does not need a longer beat split just to keep one movement each. Treat that as guidance from a platform that runs the model, and test it.

References: call them by name. Seedance 2.0 takes text, images, audio and video, with up to 9 images, 3 video clips and 3 audio clips. Seedance 2.5 takes up to 50 references in all: 30 images, 10 video clips and 10 audio clips. In the prompt you refer to uploaded files as @Image1, @Video1 and @Audio1; Runway’s guide writes them with a space, such as @Image 1, so follow whatever your platform’s picker inserts. Dreamina’s example reads, in its words, “Create a smooth video using @Image1 as the main reference.” State what each reference is for: the product to keep, the motion to copy, the voice to match. For a character, Runway’s 2.5 guide suggests naming one primary reference and repeating its key traits at each major beat instead of declaring them once. An unlabelled reference leaves the model to guess, which is our advice and not a documented rule.

Sound is part of the generation. Both versions produce audio with the picture. ByteDance describes dual-channel audio for 2.0, and Dreamina describes audio-driven lip-sync and timing from narration, dialogue or music. For dialogue in 2.5, Runway’s guide recommends plain labelled lines such as Dialogue (elderly woman): "Fly safe, my child.", and says audio can be switched off on its platform. One practical note: where Pollo’s API exposes a generateAudio switch for 2.0, its docs show it off by default, so check that audio is on for the route you use.

Iterate one variable at a time. Dreamina’s advice is to keep the prompt identical except for the one thing you are changing.

Negatives: use them sparingly. We found no negative prompt field in ByteDance’s posts, Dreamina’s pages or Pollo’s 2.0 API docs. Dreamina’s prompt guide tells you to use negatives “carefully”. Runway’s 2.5 guide narrows it further: keep negative phrasing for subtitles and music (“No subtitles. No background music, environmental sounds only.”) and describe the intended picture positively. The builder follows this by writing negatives inline, as a short closing “Negative prompt: …” sentence. Dreamina also advises skipping vague adjectives and using concrete verbs, such as “twists the cap” or “pours slowly”.

Clip length and shape. Seedance 2.0 generates 4 to 15 seconds. Seedance 2.5 generates up to 30 seconds: Dreamina’s create form (checked 2026-10-04) has a duration slider that goes to 30 s, and PixVerse and Runway list 4 to 30 seconds. With Seedance 2.5 selected, that form offers exactly six aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. The same six are the ones Pollo lists for 2.0, with an adaptive option. Resolution varies by platform and variant: Dreamina’s 2.5 form offers 480P, 720P and 1080P, with no 4K option even though Dreamina’s marketing page mentions 4K, so do not plan around 4K. On PixVerse, 2.0 Fast is limited to 480p and 720p, and 2.5 offers 480p, 720p or 1080p. Dreamina’s model menu also has a separate “Seedance 2.5 (for preview)” variant that makes 480p previews, useful for testing an idea before you spend credits on a full render.

Two restrictions to plan around. At the 2.0 launch, CapCut restricted making videos from images or videos that contain real faces; check what your platform allows if your brief animates a photo of a real person. And 2.5 is a paid feature on CapCut, PixVerse and Runway, so if you cannot see it, your plan or region is the likely reason.

A before and after, in the builder’s structure

What most people type:

A girl drinking a soda on a roof, cinematic, 6 seconds.

What the builder assembles from a brief of product, hook, scene and proof:

Create a 6-second 9:16 TikTok video ad for a sparkling yuzu soda,
optimized for Seedance 2.5. Objective: win first purchases from 18-to-30
year olds who want a less sweet drink. The opening line is: "Finally, a
soda that isn't sugar." Scene: a rooftop at golden hour, a woman in a
denim jacket cracking open a pale-yellow can. Camera and look: handheld
medium shot that pushes in to the can, warm backlight, a hint of lens
flare. Show the proof or benefit: the fizz rising and the label reading
"Yuzu" clearly, through visible action, not text overlays. Audio: she
says the line to camera; the can opening, fizz, distant city ambience.
Brand constraints: the can stays pale yellow, label legible, no health
claims. Final CTA beat: she raises the can toward the lens. Negative
prompt: on-screen captions, watermark, warped label.

That is about 130 words, above Dreamina’s recommended range for Seedance 2.0. Below is the same brief trimmed by hand to the details that define the shot (about 75 words). This is a hand-written prompt in Seedance’s own style, not builder output.

A woman in a denim jacket on a rooftop at golden hour cracks open a
pale-yellow can of yuzu soda and says to camera, "Finally, a soda that
isn't sugar." Handheld medium shot pushing in to the can, warm
backlight, fizz rising. Audio: her line, the can opening, fizz, distant
city ambience. The can stays pale yellow with a legible "Yuzu" label.
She raises the can toward the lens. No on-screen captions.

For Seedance 2.5, with a product photo attached, here is the same brief stretched to 12 seconds and laid out the way Runway’s guide describes (our adaptation of that structure). Again, a hand-written prompt in Seedance’s own style, not builder output. The builder does not write timelines or asset lists; you add those yourself on top of its text.

@Image1 is the product: the pale-yellow can with the "Yuzu" label. Keep
its colour and label unchanged.
A woman in a denim jacket opens the can on a rooftop at golden hour and
toasts the camera, handheld, warm backlight.
0-4s: Medium shot, she cracks the can open, fizz rising, ambient city
sound. Dialogue (woman): "Finally, a soda that isn't sugar."
4-8s: Slow push-in to the can in her hand, condensation catching the
light, fizz audible.
8-12s: She raises the can toward the lens and holds it, can centred.
Consistency: same woman, same jacket, can stays pale yellow throughout.
No subtitles. No background music, environmental sounds only.

These are prompt structures, not tested results. No Seedance output has been verified for any of them, and we show no video for them. Only the first block shows what the builder assembles. The trimmed version and the timeline version are our judgement of how to apply the vendors’ published advice, not measured improvements.

Seedance compared with Veo

The practical difference is inputs and length. Seedance lets you steer with video and audio, not only images, and 2.5 can hold a 30-second sequence, so a brief that is really “copy this move and this voice onto my product” suits it. Veo’s prompting is more about describing a single shot in detail, and its limits on length and shape are tighter. The structure of your brief carries across both; what changes is how much you rely on references versus words. The Veo Prompt Builder uses the same inputs with Veo’s limits, and the Veo prompt library shows Veo prompts by use case. Our Veo image-to-video prompts page covers the describe-the-motion principle that also applies when you attach a Seedance image reference.

Where to run the prompt

Official first. Dreamina is ByteDance’s own creative platform and runs Seedance 2.5, 2.0 and 2.0 Fast, plus a “Seedance 2.5 (for preview)” variant that makes 480p test renders (create form checked 2026-10-04). CapCut has Seedance 2.0 in most markets including the US since April 7, 2026, with 2.5 rolling out to paid subscribers in Europe, Asia, the Middle East and South America. Through other platforms: Runway offers Seedance 2.5 on its web app and lists 2.5, 2.0, 2.0 Fast and 2.0 Mini in its API; PixVerse offers 2.5 and 2.0 to Pro, Premium and Ultra members; Pollo AI lists Seedance 2.0 in its API documentation; and developers can use the BytePlus ModelArk API, which documents Seedance 2.0. In China the model runs in the Jimeng AI and Doubao apps. Availability depends on your region and plan, so check what your account shows. If you want to try the brief on another model afterwards, the Kling builder and the Grok Imagine builder accept the same inputs.