Seedance Prompt Examples by Use: Vertical Drama, Hooks, Motion Reference, Transitions, Product Clips

TWno1 Editorial · Published 2026-10-05

Credits and specs follow TWno1's live billing settings. Last checked: 2026-10-05

BytePlus's Seedance guide, checked in October 2026, caps the recommended prompt at 1,000 English words or 500 Chinese characters.

Past that, the model starts skipping details. Most people never hit the cap. They hit the softer version of it: a prompt so packed that the one action you cared about gets dropped. That is why every example below is short. Swap the names, places and props and they are ready to run. After each one you get what the BytePlus Seedance 2.0 prompt guide says about it and where this kind of shot usually breaks.

One caveat up front. These are worked examples built from the official guidance, not guaranteed recipes. Video models give a different take every run. Plan for re-rolls.

Five uses at a glance

Use	The one rule	Typical failure
Vertical drama shot	One person, one action, one camera move	Dialogue burned into the frame as subtitles
Opening hook	A small, clear oddity, not a big stunt	Large motion breaks bodies and hands
Motion reference	Name assets as "Video 1", never by asset ID	Too many references, no clear priority
Multi-shot transition	Number the shots; skip exact timestamps	Jumps or rollback at the seam
Product clip	Product image first; only common words on screen	Logos or garbled text nobody asked for

1. Vertical drama shot: a line of dialogue

Scene: office corridor at dusk, city lights behind the window.
Subject: Mara, a woman in her thirties, charcoal blazer, low bun.
Action: she stops, slowly turns her head to look behind her, right hand tightening on a folder.
Camera: medium close-up, locked camera.
Dialogue: Mara says quietly, "What did you just say?"
Constraints: no subtitles, no text, no logo, no watermark.

What the guide says: lock in who is doing what first, then where and the mood, then how it is shot, then style and constraints. Keep actions slow, small and continuous, and name the body part and how far it moves. Show emotion through physical detail instead of words like "furious." The API page says to put dialogue in double quotes; the 2.0 guide's examples use curly braces. Both are documented. Test both.

The usual failure is subtitles. BytePlus admits it cannot reliably stop Seedance from adding them. "No subtitles" only lowers the odds, and vertical output gets more stray subtitles than horizontal. We saw a harsher version in our own review on a different video model: once the dialogue text was in the prompt, it kept getting burned into the frame, and 11 different phrasings did not stop it. So put real subtitles in during the edit. If burned text keeps coming back, generate the picture without the dialogue in the prompt and add the voice separately.

2. Opening hook: make them stop scrolling

Scene: a home kitchen late at night, lit only by the open fridge.
Subject: Owen, a man around forty, white shirt with rolled sleeves.
Action: he opens the fridge and freezes, staring at a red envelope that should not be there.
Camera: over-the-shoulder from behind him, slow push-in toward the envelope.
Constraints: normal speed, no slow motion, no subtitles.

Short-drama writers talk about a "3-second hook." That rule of thumb comes from a single industry source we have not verified independently. What the BytePlus guide does say is that Seedance prefers slow, small, continuous actions, and that sprints and big jumps are less stable.

So skip explosions, chases and falls as openers. They are exactly the motions the guide flags. An object that should not be there, or a movement that stops dead, gives the model something it can do well and gives the viewer something to read instantly. Also watch the speed. In one 15-shot review, 4 shots were flagged as too slow under our trial threshold. "Normal speed, no slow motion" costs nothing to add.

3. Motion reference: let a clip define the action

References: Image 1 is Mara's face close-up, Image 2 is her full outfit, Video 1 is a turn-and-raise-hand motion reference.
Scene: Mara stands at an apartment door, turns as in Video 1, and lifts her hand to press the doorbell.
Camera: full-body medium shot, locked camera.
Constraints: only one person, no subtitles.

What the guide says: refer to assets by their order in the request, as "Image 1," "Video 2," "Audio 1," not by asset ID. In edit or extend tasks, call the clip "Video N" and drop the word "reference." The recommended mix is four or five assets: one or two character images (face close-up plus full body), one scene, one camera-motion video, one audio clip.

The cap and the recommendation are different numbers. Seedance 2.5 accepts up to 30 images, 10 videos and 10 audio clips, but a full load leaves the model guessing what matters. BytePlus also warns that more than four reference people lowers stability, and multi-angle character sheets can make one character appear twice.

Two hard limits. First-frame, first-and-last-frame and omni-reference modes are mutually exclusive, one per request. And Seedance 2.5 refuses reference images or videos containing real human faces, so your own photo or an actor's will be rejected. Requests with reference video also carry minimum token floors on BytePlus, so they cost more than text alone.

4. Multi-shot transition: several shots in one clip

Shot 1: café window seat, Mara looks down at her phone, locked camera.
Shot 2: close-up of the phone screen, an unread message lights up.
Shot 3: Mara looks up toward the door, slow push-in on her face.
Style: realistic, natural light, no subtitles.

What the guide says: write complex videos as numbered shots. Exact timing like "0-3 s" is unstable, so do not over-constrain each shot's length. Give each shot one camera move; mixing push, pull, pan and track raises instability.

Our own experience pulls slightly the other way, and it is worth being clear about that. On a different video model, one fix for a camera that kept moving on its own was writing the order of events with times attached ("0-2 s: she looks down; 2-3.5 s: she looks up"). That is the opposite of the official wording quoted above: BytePlus says exact seconds are unstable on Seedance. Put together: spell out the order, and do not expect the seconds to be honored.

Seams are the other failure. When you stitch or extend clips, BytePlus's guide admits the join can jump or roll back. Its workaround is to trim 6 frames from the end of clip A and 1 frame from the start of clip B in an editor. Repeated extensions also degrade quality and smear faces. The rule we used on our demo drama is stricter: chain from the previous clip's last frame only when one continuous action was split across clips. A new angle, a new location or a time jump should start fresh, because chaining carries the previous clip's defects forward.

5. Product clip: get the product and the words right

References: Image 1 is the product photo of a ceramic travel mug.
Scene: white tabletop, soft side light, the mug from Image 1 centered.
Action: the mug slowly rotates half a turn, light catching the glaze.
Camera: close-up, locked camera, shallow depth of field.
On-screen text: "Hot all day" at the bottom of the frame.
Constraints: no other logos, no watermark, no other text.

What the guide says: Seedance 2.0 can render common words, slogans, subtitles and speech bubbles, but you should avoid rare characters and symbols. Put the most important asset first.

Text is where product clips break. The BytePlus guide notes that logos and watermarks can appear without being asked for, and that a prompted countdown animation may jump around; it suggests defining such effects with a reference video instead. Our advice: any text a viewer must read goes in during the edit. Treat generated text as a layout sketch and proof it word by word.

Four prompt mistakes from our own reviews

These come from our shot-by-shot review of a 15-shot episode and an audit of 244 storyboard frames. Not every shot was made with Seedance, but none of these mistakes depends on the model.

Film-school prose. "The camera tilts up along the folds of her coat" became a video of someone stuffing paper into a shirt. Write the storyboard literally, one action per shot.
Banning names. After a rule against character names, prompts produced sentences like "a figure lifts a figure." Use "the woman," "the man" or the character's name. BytePlus also recommends naming the subject.
Examples become defaults. A few states listed as illustrations got picked up as the default, and the first 4 frames were all wrong.
Words with two meanings. A slight bow that got the word "salute" added was drawn as a raised-hand salute. A letter renamed as a newspaper clipping became the wrong prop. Keep nouns exactly as the script has them.

One more, about negatives. The Seedance guide endorses constraint lines such as "no subtitles, no logo, no watermark." Runway's Gen-4 guide says the opposite for its models: negative phrasing is not supported and can backfire. In our logs, merely writing "glasses," even as "not wearing glasses," drew glasses. Our takeaway: use constraint lines for whole categories like subtitles and logos. If a specific object must not appear, do not name it at all.

Limits: where these examples are not a good fit

Chinese script support is not spelled out. BytePlus lists Chinese among supported languages, but the pages we read do not say whether Traditional and Simplified behave the same. If you prompt in Chinese, test both for important shots.
Mispronounced Chinese dialogue. BytePlus says polyphonic or rare characters can be misread and suggests swapping in common homophones.
No real people. References with real faces are refused. TWno1's content policy also bans using a real person's photo, name or voice for sexual or nude content, and all characters must be adult fictional characters.
No repeatable seed. BytePlus lists the seed parameter for 1.5 pro and the 1.0 series, not for 2.0 or 2.5. The same prompt can give quite different results each time.
Not a TWno1 manual. The five templates follow BytePlus's own format and are general prompt craft. What TWno1 itself offers is described on Studio and the pricing page.

FAQ

How long should a Seedance prompt be?

BytePlus Seedance documentation (checked 2026-10-05) recommends no more than 1,000 English words or 500 Chinese characters, because the model skips details in longer prompts. In practice, four to six lines covering subject, action, camera and constraints are enough for one shot.

Can I write "no subtitles" in a Seedance prompt?

Yes. The BytePlus Seedance 2.0 prompt guide recommends constraint lines like "no subtitles, no logo, no watermark." BytePlus also says it cannot reliably prevent subtitles, only reduce them, and that vertical videos show more stray subtitles than horizontal ones. Add real subtitles in post.

How do I reference images and videos in a Seedance prompt?

Per BytePlus docs, Seedance references are named by their order in the request, such as "Image 1," "Video 1" and "Audio 1," not by asset ID. BytePlus recommends about four or five assets per request, warns that more than four reference people lowers stability, and refuses references that contain real human faces.

Can one Seedance clip contain several shots?

Yes. The BytePlus Seedance 2.0 prompt guide recommends writing complex videos as numbered shots (Shot 1, Shot 2, Shot 3) with one camera move each. It also says exact timing such as "0-3 s" is unstable, so avoid over-constraining each shot's length.

Why does the camera still move when my prompt says it should not?

Video models often add camera motion on their own. In our review of a 15-shot episode, the camera moved in at least 12 shots, and 11 of those prompts explicitly asked for a static camera. What helped was opening with "locked camera, no push-in, no pull-out, no pan," one action per shot, and cutting static or narration-only shots to about 3 seconds.

References

BytePlus ModelArk: Seedance 2.0 series prompt guide (checked 2026-10-05)
BytePlus ModelArk: create a video generation task (API) (checked 2026-10-05)
BytePlus ModelArk: Seedance 2.0 series tutorial (checked 2026-10-05)
BytePlus ModelArk: Seedance 2.5 (checked 2026-10-05)
Runway: Gen-4 video prompting guide (checked 2026-10-05)