AI Film Camp · Prompting
How to Format Prompts for Seedance 2.5
Treat the prompt as a shooting brief, not a caption.
Seedance 2.5 is ByteDance’s video model for single-pass clips up to 30 seconds, with synced audio and a large multimodal reference budget: up to 30 images, 10 videos, and 10 audio files in one generation. The model follows structured direction better than mood words. Treat the prompt as a shooting brief, not a caption.
That shift is what BytePlus documented in the official guide, and it is what working creators keep repeating on X: timestamps, assigned reference jobs, one event per beat, and an explicit end state.
Official sources
Start here. Everything else is commentary.
- BytePlus ModelArk prompt guide (canonical English): docs.byteplus.com/en/docs/ModelArk/2607689
- English Feishu copy: bytedance.larkoffice.com/docx/A88jd0B47oAd8zxWp5ycZFMfnxh
- Chinese prompt guide: bytedance.larkoffice.com/docx/OsiUdR1OxoDqvnxsK8LczYx7nPd
- Official handbook: bytedance.larkoffice.com/wiki/NjnWwvf4BiFYFLk2RzrcEgaunGf
- Seed team launch post, with official example prompts: seed.bytedance.com/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5
BytePlus also ships a prompt optimizer skill, sd25-pe. Install it, then run /sd25-pe on a draft. That is the closest thing to an official formatter.
Lena (@lena_z01), who previously led mobile at Dreamina, posted the official-guide announcement and the practical pitch: stop burning credits on unstructured tests. Use the BytePlus guide for timing control, reference consistency, subtitle/BGM removal, and extension. Lena
What the model is actually doing
Seedance 2.5 classifies the job from the prompt and the uploaded assets:
| Task | Trigger | What locks |
|---|---|---|
| Text-to-video | No assets | Nothing; you set ratio and 4–30s |
| Reference-to-video | @Image / @Video / @Audio as semantic refs | Nothing locked; assets steer, they do not pin frames |
| Editing | Verbs like replace, remove, modify | Source ratio and approximate duration |
| Extension | extend, continue | Source ratio |
| First / last frame | First-frame and optional last-frame images | First image’s ratio |
If an R2V prompt accidentally reads like an edit, the API can classify it as an edit and reject parameters. Pick the task first, then write the prose.
Limits vs. what is stable:
- Images: max 30, BytePlus says 1–8 subjects is usually more stable
- Videos: max 10, combined 30s; 5–10s clips work better
- Audio: max 10, combined 30s
- Storyboards: 15 panels or fewer, simple line art, little text
Fifty slots exist. Filling them all usually makes the result worse.
The six-part formula
Practice guides compress the official advice into six slots:
SubjectWho or what + Action or EventOne thing happens + Scene and EnvironmentWhere + Visual StyleStock, grain, color + Camera Movement or CutOne move + AudioOr forbid it
Only subject and action are required. Empty optional slots are better than vague adjectives. “Cinematic” is a shrug. “35mm film look, visible grain, warm faded color, soft halation” is a decision.
A short filled example:
A ceramicist in a linen apron lifts a finished bowl from the wheel and turns it slowly in the light, in a cluttered studio at golden hour with clay dust in the air, warm 35mm film look with soft halation, slow push in that settles on her hands, (sparse piano) <wheel slowing, clay scraping>.
Two or three clear sentences beat a paragraph of mood words.
MAX (@maxxmalist) popularized a close cousin of this on X and it still holds. MAX
- global scene + visual style
- character descriptions
- timestamped storyboard
- camera movement
- negative prompts
- ending constraints
The official four-layer structure
BytePlus does not require labeled headings. It does recommend this order:
- One-sentence summary — subject, location, event, genre/style, camera.
- Asset bindings — every uploaded file gets a job.
- Detailed plot — timestamps or Shot N, with action, camera, dialogue, and sound.
- Closing constraints — identity, wardrobe, set, lighting, audio negatives.
Clarity and non-contradiction matter more than heading names.
Official panda example, condensed from the BytePlus guide:
Realistic nature documentary style, natural lighting and shadows.
On a warm afternoon, on a grassy slope in the forest, a chubby panda cub rolls down the hill.
The panda has fluffy, realistic black-and-white fur and clumsy movements.
Low-angle medium-wide shot, slight handheld feel, panda kept in frame.
0s-3s: The cub lies on the slope and begins to roll sideways. Sunlight filters from the upper left.
3s-8s: It rolls toward the lower right, stops on its belly, turns its face toward the camera, and makes a soft humming sound.
Natural environmental audio only: wind, rustling grass, the soft plop of the roll.
No BGM. No subtitles.That is the official shape: summary, subject/set, timed beats, audio, negatives.
Bind every reference, or the model will invent a job
Official launch prompts use @Image 1, @Video 1, @Audio 1, and even @Clay Render 1. Do not write “use the uploaded images.” Assign a role and an exclusion.
@Image 1 defines the woman's face, hair, and green jacket only. Do not copy the background.
@Image 2 defines the coffee shop, the window, and the morning light.
@Video 1 defines camera pace and blocking only. Do not copy its characters or location.
@Audio 1 defines music rhythm only.Rules that matter:
- Number assets by upload order.
- One job per file. Face is not wardrobe is not room is not camera.
- If only part of a file should be used, name that part.
- Do not burn names onto the image. Bind names in the prompt. Labels inside a still cause duplicate characters.
- Put the identity you most need preserved first.
Preview (@previewio) put the same idea more bluntly: once the refs exist, describe less. Extra text makes the model trust the paragraph instead of the picture. Preview
FoldingEnd A.I. (@EndFolding79421) posted a live example of that discipline: one storyboard image, two character refs, an explicit “exactly one Grace,” timed beats, and a hard rule that Zack cannot appear before 00:11. That is what high-adherence Seedance prompting looks like in the wild. FoldingEnd A.I.
Write time, not a paragraph
A 12-second clip can survive one continuous description. A 30-second clip written that way wanders, rushes the middle, or freezes the opening. Official examples use second-level timestamps or Shot N.
0-8s: wide establish, first action, first camera move. End state: [frame].
8-21s: one development. One camera change. End state: [frame].
21-30s: payoff framing. End state: [final composition].Rules:
- One job per beat. One location change, one emotional turn, or one camera move.
- Every beat lands on a visible end state. The next beat inherits that frame.
- Size the window to the action. Do not cram a multi-step gesture into two seconds.
- Avoid high-frequency timestamp spam (“shake three times per second”).
- Duration and aspect ratio belong in the generation settings, not in the prompt.
Courtois Nicolas pointed to Higgsfield’s tested library as the same idea in production language: global style first, then starting positions, then shot-by-shot action, optics, physics, sound, and continuity. Courtois · Higgsfield guide
taruma sakti (@tarumainfo) pushes this further with an “auteur script”: declared camera setups persist until you change them, and state-in / state-out conditions keep continuity when you skip across beats. taruma sakti
Audio and on-screen text
Official BytePlus examples use plain English:
Dialogue (elderly woman): "Fly safe, my child."
No BGM; generate only environmental sounds and action sounds.
No subtitles.Community and practice guides also use a four-channel bracket system. It shows up in many working prompts, but it is not clearly specified as a parser on the BytePlus page. Use it if your platform likes it; fall back to plain English if results get sloppy.
| Syntax | Channel | Example |
|---|---|---|
( ) | Music / ambient bed | (low cello drone) |
< > | Sound effects | <door latch, rain on glass> |
{ } | Spoken dialogue | {We should go.} |
【 】 | On-screen text | 【Chapter One】 |
Dialogue tips from official and community practice:
- Keep lines short. Long multi-person speeches hurt lip sync and stability.
- Declare language and delivery before the line:
She says softly in Japanese: {もう大丈夫です} - Dialogue and subtitles are separate. Repeat the line in both if you want both.
- If words appear on screen, write them exactly. The model invents text when you leave it open.
Raph Guilhem (@raph_guilhem) published a reusable system prompt that encodes this: timecoded beats, end states, {dialogue}, layered audio, and a closing consistency lock. Useful if you want an LLM to draft Seedance prompts instead of writing them by hand. Raph Guilhem
Camera language
Seedance understands ordinary film terms. Pair the move with a speed word.
- Slow dolly-in / push-in
- Pull out / crane up
- Lateral track / orbit / 360
- Handheld, slight breathing
- Locked-off / static after the move completes
- Whip pan, smash cut, hard cut — only if you want a cut
For one-takes, say so: single continuous take, smooth camera movement, no cuts. Official concert and Peking-opera demos are written that way.
Editing and extension
Use the verb that matches the job.
Edit
Edit @Video 1. Keep characters, actions, and visual style unchanged.
Adjust only the camera movement.
0-4s: micro-FPV past the pan...Extend
Extend the video. Continue from the visuals and subjects in @Video 1
and generate another clip, keeping character, scene, style, and sound consistent.Framer (@Framer_X) showed the practical version of this: generate 10 seconds, re-upload that clip as the video reference, drop the stills, then prompt the next 15 seconds as SHOT 1… SHOT 7 with locked character names and short dialogue. That is how people get multi-minute scenes without restating the whole world every time. Framer
The workflow that keeps winning on X
The highest-engagement Seedance 2.5 threads are not “write a longer prompt.” They are:
- Design cast and boards in an image model.
- Bind those stills as
@Imageroles. - Direct Seedance with a shot list.
el.cine (@EHuanglu) has the two most-circulated versions of that pipeline: ChatGPT Image storyboard + cast sheets, then Seedance 2.5 for the scene. el.cine · second thread
jboogx.creative (@jboogx_creative) made the same case as a cost argument: Seedance 2.5 is strong and expensive, so storyboard first and cut regenerations. He also posted the official Feishu guide early. jboogx.creative · Feishu guide
Copy-paste skeleton
Paste into your AI chat
Short one-take variant, adapted from official launch demos:
Paste into your AI chat
Best written guides after the official docs
- Segmind six-part formula and syntax reference
- Rendley walkthrough of the BytePlus page, including the audio-syntax caveat
- Tryonr official-formula recap
- Oakgen 30-second production-brief skeleton
- Official launch prompts, transcribed
- Magnific: 10 tested prompts
- Together AI quickstart
- Community archives: awesome-seedance-2x-prompts, awesome_seedance_2_5
The only rules that keep showing up
Name one subject specifically. Give that subject one action per beat. Tell the camera what to do. Bind every reference to a job and an exclusion. Write the 30 seconds as time, not as a vibe. Describe the sound, or explicitly forbid music and subtitles. Then stop adding adjectives.
If you only open two links, open the BytePlus guide and the Seed launch examples. If you only follow two people on X for this, follow Lena’s official-guide thread and el.cine’s storyboard-to-Seedance workflow.