---
name: seedance-2-0-reference-prompts
description: Format, rewrite, expand, or audit Dreamina Seedance 2.0 series video prompts that use ordered image, video, or audio references. Use for multimodal reference-to-video, subject binding, shot-sequenced narratives, native dialogue and sound, video editing, forward or backward extension, track completion, First/Last Frame planning, multi-character consistency, text generation, and Seedance 2.0 troubleshooting.
---

# Format Seedance 2.0 Reference Prompts

Turn ordered assets and a scene brief into one concise, copy-ready Seedance 2.0 prompt. Treat the model as a multimodal director: references establish spatial truth, while shot order establishes temporal change.

## Load the exact rules

- Read [references/seedance-2-0-spec.md](references/seedance-2-0-spec.md) before formatting or auditing any prompt with references. Use it for canonical asset labels, limits, task modes, subject binding, shot structure, audio/text symbols, and validation.
- Read [references/prompt-patterns.md](references/prompt-patterns.md) when the request involves several assets, more than one character or shot, editing, extension, track completion, First/Last Frame, dialogue, text generation, or a request for a template or worked example.

## 1. Establish the task contract

Classify the request before writing:

- **Text-to-video**: Generate from a scene description without reference media.
- **Image-to-video**: Animate one first frame or bridge separate first and last frames.
- **Multimodal reference**: Extract identity, scene, motion, camera, style, effects, voice, or rhythm from ordered assets to create a new video.
- **Video edit**: Add, remove, or modify a bounded element in an existing video.
- **Video extension**: Generate content before or after an existing video.
- **Track completion**: Join two or three ordered videos with described transitions.
- **Combined task**: Reference one asset while editing another.

Ask only when a missing fact changes asset order, subject identity, the source video, edit scope, extension direction, dialogue language, or intended ending. Infer harmless production details and state them briefly after the prompt when useful.

## 2. Build an internal asset registry

List assets in upload order and map each to one job. Normalize labels exactly as:

```text
Image 1
Video 1
Audio 1
```

Do not add an `@` prefix. Do not collapse the label to `Image1`. Seedance 2.0 and Seedance 2.5 use different conventions.

For each asset record:

1. Named subject or production role.
2. Attributes or behavior to inherit.
3. Incidental content to ignore.

Map every character, prop, scene, motion source, camera source, and voice separately. Never rely on “respectively.” For simple repeated subject bindings, `Courier@Image 1` is valid. For complex scenes, define a stable role once and use that same role in every shot.

## 3. Choose the correct temporal structure

- Use natural prose for one subject performing one simple action in one shot.
- Use `Shot 1`, `Shot 2`, and `Shot 3` for multi-action or multi-camera clips.
- Give each shot one primary camera movement and a small number of compatible actions.
- When a referenced camera move must finish on a focal subject, direct that same move to settle on the subject. Do not stack a new push, tilt, orbit, or crane move onto its ending; use a cut when a second camera idea is essential.
- Prefer chronological shot order over exact second ranges. Seedance 2.0 timing responds more reliably to ordered shots than frame-like timestamp control.
- Express action through body part, direction, amplitude, speed, and force.
- Describe the physical transition between consecutive actions so motion retains inertia.
- Replace abstract emotion with visible expression and posture.

## 4. Compose in production order

Use this order unless a mode-specific pattern requires otherwise:

1. `[Asset Roles]`
2. `[Generation Goal]`
3. `Shot 1`, `Shot 2`, and later shots in chronological order
4. `[Global Direction]` for visual style, lighting, image quality, and sound bed
5. `[Constraints]` for likely failure modes only

Within a shot, write:

```text
Camera: <shot size plus one movement or cut>.
Subject action: <specific body movement, expression, and continuity>.
Space: <position or spatial change>.
Audio: <dialogue, ambience, effect, or music>.
End state: <visible state inherited by the next shot when needed>.
```

Omit empty fields. Keep the most important identity asset first. Do not bury essential bindings under cinematic adjectives.

## 5. Apply the mode-specific contract

- **Reference generation**: Write `Reference <dimension> from Image 1 / Video 1 / Audio 1 to generate...` and name what must stay consistent.
- **Editing**: Directly name the editing source: `Strictly edit Video 1.` Do not write `reference Video 1`, which can change task interpretation. Bound the edit and list preserved content.
- **Forward extension**: Write `Extend Video 1 forward.` Continue its final pose, frame composition, motion direction, lighting, audio character, and story state.
- **Backward extension**: Write `Extend Video 1 backward.` End the new segment on the original first-frame state.
- **Track completion**: Order the sources and transitions explicitly. Use no more than three source videos.
- **Combined task**: Separate authority: `Reference <dimension> from Image 1; strictly edit Video 1; <specific change>.`
- **First/Last Frame**: Use the dedicated mode for exact boundary matching. In multimodal reference mode, a prompt can request first/last-frame use, but the boundary is less strict.

## 6. Format sound and generated text

Use these content markers when they improve clarity:

```text
（Music direction）
<Sound effect>
{Dialogue}
【Subtitle or on-screen text】
```

Keep dialogue in one language except proper nouns. State the language before dialogue whenever pronunciation or language choice may be ambiguous. When matching a voice, combine `Audio 1` with a short description of pitch, texture, pace, emotion, and delivery.

## 7. Validate before returning

Check all of the following:

- Every used asset label maps to one named job.
- Labels use `Image 1`, `Video 1`, and `Audio 1` with title case, a space, and no `@` prefix.
- Every subject keeps one stable role name across shots.
- Character count, identity, clothing, props, and screen position cannot swap or duplicate.
- Each shot uses one main camera movement and physically coherent actions.
- Precise timestamps are omitted unless the user explicitly requires them.
- Edit and extension prompts name `Video N` directly rather than calling it a reference.
- Editing specifies exactly what changes and what remains untouched.
- Extension specifies the boundary pose, composition, motion, lighting, and audio continuity.
- Dialogue language is consistent and special content uses the correct marker.
- The prompt explicitly prevents likely subtitles, logos, watermarks, duplicate characters, or style drift when relevant.
- UI/API parameters are omitted from prompt prose unless the user explicitly asks for an operational parameter.

## Output contract

Return:

1. A short `Asset map` when two or more references are used.
2. One final prompt in a single plain-text code block, ready to paste.
3. A short `Assumptions` or `Watch-outs` note only when it materially helps.

Do not provide competing variants unless requested. Do not expose internal chain-of-thought or a lengthy explanation of the formatting process.