LESSON 7 · REFERENCES

Multimodal references and reference syntax

Seedance 2.0 can rely on more than just text. You can feed images, video and audio into the generation — and each reference type owns its own «zone of responsibility». In this lesson we'll break down who's responsible for what, and how to use tags like [Image 1], [Video 1] to point the model precisely at which reference your instruction refers to.

Four sources of information

The model accepts four types of input. The key is not to dump everything into one and the same type, but to distribute the tasks: text is good for one thing, video for another. The cleaner the division of roles, the more predictable the result.

📝
TEXT

Space

Appearance, mood, style, environment, light. Everything about «how it looks» and «what atmosphere».

🖼️
IMAGE

Look and pose

Up to 9 images: a specific character, scene, style, reference pose or frame composition.

🎬
VIDEO

Time

Up to 3 clips (≤15 s): timing, rhythm of gestures, trajectory and camera movement — everything that unfolds over time.

🎵
AUDIO

Rhythm and sound

Up to 3 tracks (≤15 s): sets the rhythm of action, the pace of editing, the sonic backdrop of the scene.

🔑 The core principle
Text — for the spatial (how it looks, what mood, what style). Video reference — for the temporal (how it moves, in what rhythm, with what camera dynamics). Don't describe complex motion with a stack of adverbs («quickly, smoothly, gracefully, suddenly…») — give a video reference, and the model will copy the dynamics more accurately than any words.

Universal reference syntax

Uploading references isn't enough — the model needs to understand which of them a particular phrase in the prompt refers to. For this there's a single tagging syntax: you reference each source right in the prompt text by its number.

[Image 1]

The first uploaded image. For example — the hero's appearance.

[Image 2]

The second image. For example — a reference pose or composition.

[Video 1]

The first video reference. The source of motion and timing.

[Audio 1]

The first audio track. Sets the rhythm and sound.

Numbering follows the order of references within each type: the first image is [Image 1], the second is [Image 2], and so on, independently of video and audio, which have their own running numbering ([Video 1], [Audio 1]).

The reference-operation formula

For a tag to work, frame the instruction as a meaningful operation: what to take from the reference (Reference / Extract / Combine), from exactly which one ([Image N]), what to generate, and what to keep unchanged in the process.

Reference / Extract / Combine + [Image N] → generate [scene], maintaining consistent [subject]
1
Operation
Reference / Extract / Combine — what exactly we do with the reference
2
Tag
[Image N], [Video N], [Audio N] — where we take it from
3
Goal
generate [scene] — what we want to get as output
4
Anchor
maintaining consistent [subject] — what we keep unchanged

Example: choreography through tags

Watch how a single scene is assembled from three references. We take the heroine's appearance from one image, her pose from another, and all the emotion and motion dynamics from the video. Each instruction is tied to its own tag, and the model doesn't get confused.

The woman in [Image 1] walks up to the mirror. Her pose references [Image 2]. The emotion and motion of the breakdown should fully reference [Video 1].

Here the division of roles is crystal clear: [Image 1] — who this is (appearance, spatial), [Image 2] — how she stands (reference pose), [Video 1] — how she moves and what she feels (temporal). The text glues it all into one coherent scene.

R2V · reference → video

The golden horse from a reference

The pose and gallop motion are taken from the source reference, and the object then smoothly transforms — transfer of motion and form.

Reference limits

Each reference type has its own ceiling. Keep these numbers in mind — the model will accept exactly this much input and no more.

9
images maximum
3 · ≤15s
video (total)
3 · ≤15s
audio (total)
TypeLimitWhat it's responsible forTag
Imageup to 9character, scene, style, pose, composition[Image 1…9]
Videoup to 3, ≤15 s totaltiming, rhythm of gestures, camera movement[Video 1…3]
Audioup to 3, ≤15 s totalrhythm, tempo, sonic backdrop[Audio 1…3]
✅ Tip for practice
Don't duplicate the same task across different reference types. If the dynamics are set by [Video 1], don't also describe them with adverbs in the text — this creates a conflict of instructions. One aspect — one source: appearance in words or [Image N], motion via [Video N], rhythm via [Audio N].
multimodality[Image N][Video N][Audio N]reference / extract / combinedivision of roles

Try Seedance 2.0 yourself

Generate your first clip in ZenCreator — Seedance 2.0 is available in the Image-to-Video tool.

Start Creating
PreviousNext lesson