The first prompt I wrote for a 30-second Seedance 2.5 clip was the same prompt I would have written for a 5-second one. One subject, one action, one camera move. Thirty seconds later I had a beautiful shot of a woman standing on a rooftop doing almost nothing for half a minute. Technically perfect. Completely useless.
That is the trap nobody warns you about. Every text-to-video tutorial out there teaches you to write a tight, single-beat prompt — because until recently, that was the only length you could generate. Seedance 2.5 generates up to 30 seconds as one continuous clip, with no stitching, and a single-beat prompt simply does not have enough story in it to fill that runway.
This tutorial covers the whole loop: how text to video works end to end, how a short-clip prompt is structured, and — the part competitors skip — how a 30-second prompt is structured differently. If you want the general prompt formula on its own, the Seedance prompt guide has it. This one is about the workflow and the long-form structure.
What Actually Changed in Seedance 2.5
ByteDance previewed Seedance 2.5 on June 23, 2026 at the Volcano Engine FORCE conference in Beijing. Three of the announced changes matter directly to how you write a text prompt:
| Capability | Seedance 2.0 | Seedance 2.5 (claimed, preview) |
|---|---|---|
| Single-pass clip length | up to 15s, stitched from segments | up to 30s, native single clip |
| Multimodal references | up to 12 | up to 50 (images, video, audio) |
| Output resolution | — | native 4K |
| Editing | regenerate | targeted local edits |
ByteDance also states native audio synchronization and roughly 20% better prompt adherence. Treat all of this as preview-stage claims until you have run it yourself — the model went public in early July 2026 through CapCut and Dreamina, with enterprise beta on BytePlus and Volcano Engine since the announcement. (Community chatter puts broader API access around July 10; that is unofficial and not confirmed by ByteDance.) For the full spec breakdown, see what Seedance 2.5 is.
The headline for prompt writers is the first row. Native 30 seconds means the model plans the whole clip in one pass — so it needs a plan from you.
The Text-to-Video Workflow, End to End
Interfaces differ between CapCut, Dreamina, and third-party front ends, so here is the flow in generic terms rather than button names.
- Pick the length before you write. Length determines prompt structure, not the other way round. Decide 5–10 seconds or 20–30 seconds first.
- Write the prompt to match that length. Short = one beat. Long = a shot script. The next two sections cover both.
- Set aspect ratio and resolution. 16:9 for landscape, 9:16 for vertical, 1:1 for feed posts. Choose before generating; re-cropping a 4K clip afterward costs you framing.
- Add references if identity matters. With up to 50 multimodal inputs, a face, a product, or a location can be pinned by reference instead of described in words. Words drift; references do not.
- Generate one draft, watch it twice. Once for the subject, once for the camera. Most people only watch for the subject and miss that the camera never moved.
- Change one variable, regenerate. One line per iteration. Change three things and you learn nothing about which one worked.
- Use local editing for small misses. If 27 of 30 seconds are right, edit the region rather than rerolling the whole clip.
You can run this loop in Seedance 2.5 AI straight from the browser — no install, no API key, and the same prompt box for short tests and 30-second runs.
Short-Clip Prompts: One Beat, Four Lines
For anything under about 10 seconds, the classic structure still wins: subject, action, camera, style. One beat, one continuous motion arc, nothing more.
A barista in a denim apron pours milk into a flat white.
She lifts the pitcher and sets the cup down on the counter.
Medium close-up, slow push-in.
Warm morning window light, shallow depth of field, calm mood.That is a complete short clip. The action has a start and an end, the camera does exactly one thing, and the style sets tone without stacking adjectives. Do not add a second beat here — 8 seconds cannot hold two.
The 30-Second Prompt Is a Shot Script, Not a Sentence
Here is the increment most guides miss entirely. When you stretch that same four-line prompt across 30 seconds, the model has one beat of instruction and 30 seconds of runway. It fills the gap with drift: aimless camera float, repeated gestures, a subject that wanders out of character.
A 30-second Seedance 2.5 prompt needs three to five labeled beats with explicit time budgets. Structure it like this:
- A global header — the things that must stay constant across the whole clip: subject description, wardrobe, location, grade, aspect ratio, audio bed.
- Beat blocks — each with its own duration, action, and camera. Roughly 6–10 seconds per beat.
- A closing state — where the clip should land, so the ending is deliberate instead of a hard cut mid-motion.
The header is the part people forget, and it is what carries identity across beats. Everything you write once at the top is a constraint the model applies for the full 30 seconds. Everything you write inside a beat is local.
Copy-Paste 30-Second Template
GLOBAL: A woman in her thirties, short dark hair, olive canvas jacket.
Rooftop of a concrete building at dusk, city skyline behind her.
Consistent look throughout: teal-and-amber grade, soft film grain, 16:9.
Audio: low city hum, distant traffic, no music.
BEAT 1 (0-8s): She stands at the railing looking out over the city.
Wide establishing shot, very slow drift forward.
BEAT 2 (8-18s): She turns from the railing and walks toward camera,
pulling her jacket closed. Medium shot, slow dolly back matching her pace.
BEAT 3 (18-26s): She stops and looks off-frame left, exhales.
Medium close-up, static, slight handheld breathing.
BEAT 4 (26-30s): She looks down and half-smiles.
Close-up, holds steady. End on her face, no cut.Swap the subject and the location; keep the skeleton. Four things make it work: the global block never moves, each beat owns a time slice, every beat names its own camera behaviour, and the last beat says where to land.
Two rules from testing that are worth more than any adjective. One camera move per beat — a beat with a dolly and a pan and an orbit resolves to mush. And write physical actions, not emotions: "exhales and looks down" renders; "feels conflicted" does not.
If you want to keep a face identical across all four beats, describing it in the header is the weaker option — attach a reference image instead. Reference-led prompting is what the 50-input ceiling is really for, and the Seedance 2.5 AI video generator keeps text, image, and reference modes in the same workspace so you can switch without leaving the page.
Audio Is Now Part of the Prompt
Because Seedance 2.5 generates synchronized audio natively, silence in your prompt is not neutral — the model infers a soundscape from the visuals and sometimes guesses wrong. Name it in the global header: ambient rain on glass, no dialogue or sparse piano score, no diegetic sound. One line. It saves you a reroll.
Five Failures and the Fix for Each
| What you see | Why it happens | Fix |
|---|---|---|
| Subject drifts or changes face mid-clip | Identity lives only in beat 1 | Move the description into the GLOBAL header, or pin it with a reference image |
| Nothing happens for 10 seconds | One beat stretched over 30s | Split into 3–5 beats with time ranges |
| Camera feels random | No camera line in a beat | Give every beat exactly one named move |
| Style shifts between beats | Grade described inside a beat | Grade, grain, and palette belong in GLOBAL only |
| Clip ends mid-gesture | No closing state | Add a final short beat that says where to land |
Notice the pattern: four of the five are structural, not descriptive. Better adjectives almost never fix a long clip. Better scaffolding does.
Frequently Asked Questions
Do I have to use all 30 seconds? No, and you often should not. Test your concept at 5–8 seconds with a single beat, confirm the look and the subject, then expand the winning prompt into a beat structure. Iterating at full length is slow and expensive.
How many beats fit in 30 seconds? Three to five is the practical range, at roughly 6–10 seconds each. Six or more and each beat gets too little runway to complete a motion arc.
Does the model need me to number the beats?
Labeling helps. Explicit markers like BEAT 2 (8-18s) give the model a schedule instead of a paragraph, which is exactly what a single-pass 30-second generation needs.
Can I mix text-to-video with reference images? Yes — that is the point of the 50-input ceiling. Text carries motion and camera; references carry identity and look. Use each for what it is good at.
Is text-to-video or image-to-video better for consistency? Image-to-video anchors harder because the first frame is fixed. If a brand asset or a specific face must be exact, start from an image. If you want freedom in framing, start from text.
The Bottom Line
Seedance 2.5 text to video did not just make clips longer — it changed the unit of writing. A short clip is a sentence. A 30-second clip is a shot script with a global header, timed beats, one camera move each, and a deliberate ending. Write it that way and the drift problem mostly disappears.
Start small: run the four-line short prompt, then expand the same idea into the beat template above and compare. Both fit in one browser tab in Seedance 2.5 AI, so you can test a short version and a 30-second version back to back and see exactly where structure earns its keep. When you are ready to go long, open the text-to-video generator and paste the template.
Sources
Seedance 2.5 capability figures above — 30-second native single-pass clips (vs. 15s in 2.0), up to 50 multimodal references (vs. 12), native 4K, native audio synchronization, targeted local editing, and the ~20% prompt-adherence improvement — come from ByteDance's June 23, 2026 preview announcement at Volcano Engine FORCE, as reported by:
These are preview-stage claims from ByteDance, not independently benchmarked results. Public availability began in early July 2026 via CapCut and Dreamina, with enterprise beta on BytePlus and Volcano Engine from June 23. The commonly cited July 10 API date circulates in community channels and has not been confirmed officially. Verify current limits, pricing, and regional availability on ByteDance's official Seedance pages before relying on them.
