# Reference Workflow

For conditional motion-transfer, editorial-order and accepted-frame continuation examples, load [reference and continuity cards](continuity-example-cards.md). Their missing assets must be bound and inspected before use; conceptual states are not accepted footage.

## Reference Tag Syntax

Seedance 2.0 binds uploaded assets with an `@`-mention typed directly in the prompt: type `@` in the prompt field to pick an uploaded file, or write the tag inline. Tags are assigned by type and upload order.

- Images: `@Image1`–`@Image9` (Chinese surfaces: `@图片1`–`@图片9`)
- Videos: `@Video1`–`@Video3` (Chinese surfaces: `@视频1`–`@视频3`)
- Audio: `@Audio1`–`@Audio3` (Chinese surfaces: `@音频1`–`@音频3`)

Write each tag with the job it does, never on its own: `@Image1 as the first frame`, `follow @Video1 for camera movement`, `@Audio1 for background music`. Keep tags literal — do not translate `@Image1` into another language, renumber it, or rewrite it as a bracketed `[Image1]`; the platform's `@`-parser does not recognize the bracket form, so a mistyped tag silently fails to bind its reference.

## Asset Role Map

Before writing prompt prose, assign every uploaded asset a role. Role mapping prevents accidental transfer of identity, logos, scene ownership, or incompatible camera and motion instructions.

| Asset | Good Roles | Avoid |
|---|---|---|
| Image | identity, product, pose, costume, environment, first frame, last frame | asking it to define unseen motion |
| Video | motion, camera, pacing, blocking, timing, gesture rhythm | copying protected identity, logo, or scene ownership |
| Audio | rhythm, tempo, mood, ambience, delivery tone, music texture | assuming voice, song, or likeness authorization |
| Text brief | action, genre, camera plan, constraints | replacing concrete reference roles with vague mood words |

## Rules

- Preserve reference tags exactly.
- Give every reference one primary role before writing style language.
- Do not ask one reference to control incompatible roles unless the tradeoff is explicit.
- Use owned, licensed, public-domain, or clearly authorized references.
- Write what should transfer and what should not transfer.
- When authorization is unclear, transfer broad motion, tempo, mood, or production function rather than protected identity.
- Treat multimodal reference generation, video edit, video extend, and first/last-frame generation as separate tasks. They can share assets, but the prompt should name the active workflow.
- If audio and video references compete, make the video silent when audio timing must dominate, or state that the video controls camera/motion only and `@Audio1` controls tempo.
- In sequences, separate canonical references from accepted continuity sources: canonical identity/product references control immutable design, while accepted previous footage controls transient opening state.
- Never let a motion reference overwrite continuity locks, completed beats, reserved beats, or exact reference tags.

## Workflow-Specific Patterns

| Workflow | Use this wording | Avoid |
|---|---|---|
| Multimodal reference | `@Image1 controls product identity; @Video1 controls camera rhythm; @Audio1 controls tempo only.` | `Use all references for style.` |
| Video edit | `@Video1 is the source clip; preserve composition and timing, change only [lighting/background/VFX].` | Regenerating the whole concept from scratch. |
| Video extend | `@Video1 is the previous clip; continue the same shot for [duration] and preserve last-frame continuity.` | Starting a new scene with no continuity anchor. |
| First/last frame | `@Image1 is first frame; @Image2 is final visual target; generate the continuous transition only.` | Asking the last frame to be only "mood." |
| Audio reference | `@Audio1 controls tempo and energy; do not copy protected voice, song, or performance identity.` | Treating audio as authorization proof. |

## Role Examples

| Situation | Strong map |
|---|---|
| Product ad | `@Image1 controls product identity; @Audio1 controls tempo only.` |
| Motion transfer | `@Video1 controls side-step choreography only; do not transfer performer, costume, room, or logo.` |
| Style reference | `@Image2 controls warm bar atmosphere only; product identity remains from @Image1.` |
| First-last frame | `@Image1 is first frame; @Image2 is target end frame; transition occurs through light sweep, not product deformation.` |
| Edit/extend | `@Video1 is the source clip; preserve subject and camera path, replace only the failed lighting beat from 3s to 5s.` |

## Motion Transfer

Field-observed technique; test before promising results. Probably the most under-used reference capability: a donor video drives choreography or camera rhythm while an image keeps identity.

- Pair one donor `@Video1` with one identity anchor `@Image1`, and write the exclusion explicitly: `@Video1 controls the choreography only - nothing of its appearance, performer, costume, room, or logo transfers.`
- Pick donor clips with one clear action, a clean silhouette, and a steady camera. Busy multi-person footage transfers noise, not motion.
- Mute the donor clip before upload unless its sound should drive timing; if it keeps sound, state which reference owns the clock.
- Transfers well: choreography, gesture timing, camera rhythm, blocking. Transfers poorly: fine hand detail, multi-person sync, facial performance.
- Use only owned, licensed, stock, mocap, rehearsal, or self-recorded donor footage; real-person donors transfer general motion only, never likeness.

## Template

`@Image1 controls product identity. @Video1 controls camera pace only. @Audio1 controls tempo only. Preserve the subject from @Image1; do not copy characters, logos, music, voice, or environment from @Video1/@Audio1.`

## Sequence Transfer Template

`@Video1 is the accepted previous clip and controls only the actual observed opening state, camera phase, motion phase, ambience, and environment arrangement. @Image1 controls canonical identity. Preserve both tags exactly. Do not copy unrelated identity, costume, logo, or future action from any reference.`