Seedance 2.5 Prompt Guide: One-Take Videos with Reference Roles
Short answer: Seedance 2.5 wants a director's brief, not a keyword list. Give it a FORMAT, tell each reference exactly what it controls (@Image1 = identity only), describe the STARTING STATE, write a TIMELINE with one visible action per beat, fix the CAMERA, list what must not change, describe the AUDIO and close with CONSTRAINTS. Do that and the model holds a character, an outfit and a location for 30 seconds without a single cut.
What is new in Seedance 2.5
- 30 seconds in one take (up from 15 in Seedance 2.0), plus multi-round extension for longer scenes.
- Up to 50 references — 30 images, 10 videos, 10 audio clips — each addressed by name (
@Image1,@Video1,@Audio1) and given a role. - Timestamp-level control: second ranges in the prompt are honoured, so a beat lands where you put it.
- Native multi-shot and editing: green-screen and camera-perspective edits, reference-based re-draws of an existing clip.
- Availability: 480p/720p on fal.ai and WaveSpeed, BytePlus ModelArk (
dreamina-seedance-2-5-260628), and 1080p early access inside Higgsfield since August 2026.
The eight blocks
| Block | What to write | Example |
|---|---|---|
| FORMAT | Duration, aspect ratio, one continuous take or numbered shots | one continuous take, 20 s, 9:16 vertical |
| REFERENCE ROLES | One job per reference, plus what NOT to copy | @Image1 controls identity only — not its pose, background or lighting; @Image2 controls the outfit; @Audio1 is the music bed |
| STARTING STATE | Where the subject, camera and light are at 0 s | she stands at the café door, back to us, rain outside, camera at eye level 3 m behind |
| TIMELINE | Second ranges, one visible action per beat, where motion settles | 0-4 s she pushes the door and steps into the rain; 4-12 s she walks toward us, hood up; 12-20 s she stops, looks up, the rain thins, she settles still |
| CAMERA | One explicit move, lens, screen position | slow tracking shot backwards at eye level, 35 mm, subject centre-left |
| CONTINUITY | Invariants that must not change | red coat, hood up, umbrella in the left hand throughout; screen direction left to right |
| AUDIO | Dialogue with timing, room tone, foley, music | rain on awnings; footsteps; at 12 s she says "Finally." (soft, relieved); piano enters at 14 s |
| CONSTRAINTS | What must not happen | no cuts, no slow motion, no second person, no on-screen text or watermark |
A complete example (20 s, 9:16)
FORMAT: one continuous take, 20 s, 9:16 vertical. REFERENCE ROLES: @Image1 controls identity, face and body only — do not copy its pose, background or lighting. @Image2 controls the red wool coat and the black umbrella. @Audio1 is the music bed (soft piano), fade in at 14 s. STARTING STATE: a woman (the @Image1 identity) stands inside a café doorway with her back to us, rain falling outside, warm café light behind her, camera at eye level about 3 m behind. TIMELINE: 0-4 s — she pushes the glass door open and steps out into the rain, hood up. 4-12 s — she walks toward the camera along the wet pavement, umbrella in her left hand, reflections in the puddles; the camera tracks backwards at her pace. 12-20 s — she stops, lowers the umbrella, looks up; the rain thins to a drizzle and warm light spreads across her face; she settles still for the last second. CAMERA: slow backwards tracking shot at eye level, 35 mm, subject centre-left, no zoom. CONTINUITY: same face, hair, red coat and umbrella throughout; screen direction stays left to right; lighting shifts only from cool rain to warm daylight. AUDIO: rain on awnings and footsteps; at 12 s she says "Finally." (soft, relieved); @Audio1 piano enters at 14 s under the room tone. CONSTRAINTS: no cuts, no slow motion, no second person, no on-screen text, no watermark.
Reference roles: one job per reference
The single biggest quality lever. Untagged references get blended — the model copies the pose from your identity photo, the lighting from your outfit photo and the background from both. Tell it explicitly: @Image1 controls identity, face and body only — do not copy its pose, background or lighting. Use separate references for outfit, location, style board and audio, and never give two references the same job. Clean, front-facing canon photos work best; see how to make a reference photo.
Writing the timeline
- One visible action per beat. "She turns and smiles" is two beats.
- Cause → effect → sound → reaction, as separate sentences. "The door opens. Rain hits the pavement. She pulls the hood up."
- Say where the motion settles. Every take should end on a held pose; otherwise the last second morphs.
- Keep beats proportional. A 20-second take holds three or four beats; eight beats produce chaos.
Extension and multi-shot
For scenes longer than 30 seconds, generate the base take, then extend in rounds while repeating the REFERENCE ROLES and CONTINUITY blocks verbatim. If you actually want cuts, switch to numbered shots ("Shot 1 (6 s): …, Shot 2 (8 s): …") — or use Kling 3.0, which is built around cut sequences. GoldenPrompts' Trends atelier writes both forms and adds the entry/exit frames between scenes automatically.
Common mistakes
- Spec tokens ("4K, 60fps, 48 kHz") — noise; choose resolution in the platform.
- Negative prompts — write a CONSTRAINTS line instead.
- Ten references with no roles — the model averages them.
- Camera contradictions — "locked-off tripod" and "whip pan" in one take.
- No settle — a take that ends mid-motion drifts in the last frames.
FAQ
What is Seedance 2.5?
ByteDance's flagship video model, launched on July 31, 2026. It generates up to 30 seconds in a single continuous take with multi-round extension, accepts up to 50 references (30 images, 10 videos, 10 audio clips) with explicit roles, supports timestamp-level control, native multi-shot and reference-based editing. Seedance 2.0 (February 2026) is still sold as the cheaper tier.
How do @Image references work?
You attach references and name them in the prompt: @Image1, @Video1, @Audio1. Each reference gets one job — "@Image1 controls identity only, do not copy its pose, background or lighting" — so the model knows what to keep and what to invent. Up to 30 images, 10 videos and 10 audio clips per generation.
Where can I run Seedance 2.5?
Through API aggregators such as fal.ai and WaveSpeed (480p and 720p), on BytePlus ModelArk (model dreamina-seedance-2-5-260628) and inside Higgsfield, which added 1080p early access in August 2026. Field names differ per platform, so map the blocks to the fields you see.
What does it cost?
In September 2026: WaveSpeed lists about $0.36 per second at 720p and $0.162 per second at 480p (with a 10% promotion), plus about $0.11 per second for reference-based edits; BytePlus ModelArk bills per token, roughly $0.10 per second at 480p and $0.23 per second at 720p. A 30-second 720p take therefore costs roughly $7-11 depending on the provider.
Seedance 2.5 or Kling 3.0 or Veo 3.1?
Seedance 2.5 for long one-take stories and heavy reference control; Kling 3.0 for cut multi-shot scenes with named-speaker dialogue (3-15 seconds, up to six shots); Veo 3.1 for 8-second realism with native audio and up to three ingredient images. Many teams storyboard in Kling and shoot hero takes in Seedance.
Do negative prompts or fps tokens help?
No. Write exclusions as a CONSTRAINTS line ("no cuts, no slow motion, no duplicate subject, no on-screen text"). Spec tokens such as "4K 60fps" add nothing — the model renders at its own settings; ask the platform for the resolution instead.
Want the eight blocks written for you, scene by scene? GoldenPrompts builds Seedance 2.5 briefs — reference roles, timeline, camera, continuity, audio and constraints — from a few clicks. Free to start: 24 hours of everything, no card.