Seedance 2 · Video

Prompt GuideNEW

Seedance 2.0 natively generates audio + video together (turn on generateAudio for sound). This channel supports text-to-video and image-to-video (up to 2 reference images). This guide distills the official docs into practical prompt-writing rules.

Mental model

Seedance 2.0 is a multimodal AI director: it reads text and reference images, and internally splits them into a "spatial layer" (what is in frame) and a "temporal layer" (how it changes over time). So a good prompt is an engineering instruction, not a flowery description — who, in what scene, doing what action, how the camera moves, and in what temporal order.

Basic formulas: two generation modes

This channel offers two modes only: text-only (text-to-video) and "text + reference images" (image-to-video). Up to 2 reference images, or a first-frame / last-frame image. Video or audio inputs are NOT supported, and editing or extending an existing video is NOT supported.

① Text-to-video

Generate purely from a text description — best when you have no reference image and just an idea.

② Image-to-video

Lock subject / scene look with reference images, then describe action & camera in text. Up to 2 reference images (imageUrls), or a first / last frame (firstFrameUrl / lastFrameUrl).

Reference the <subject / scene> in <ImageN>, have it <do what>, with <camera movement>
First/last frame: open on the first-frame image, end on the last-frame image, with <transition & action> in between

⚠ Reference images must be publicly reachable; for multiple subjects, bind each to a specific image (see "Define subjects" below).

Advanced formula: 8 elements

Precise subject + action detail + scene + lighting/tone + camera + visual style + quality + constraints

First lock down "who" is "doing what", then "where" and "what mood", then tell the model "how to shoot it", and finally tighten with style, quality and constraints.

① Define subjects

Pattern: define the [2-3 stable static traits] in <ImageN> as <SubjectN> (clothing, hairstyle, look).
Multiple subjects: define each with a unique, stable label and keep reusing it (e.g. "cop", "thief").
Simple scenes without predefined subjects: bind each mention as <subject>@<ImageN>, e.g. "Zhang@Image1" (up to 2 reference images on this channel).
Keep it concise, avoid contradictory traits; express spatial relations with reference images.

② Use shot-by-shot timing

The model decouples space and time. The ideal form is a timeline of shots: describe Shot 1 / Shot 2 / Shot 3 in event order, each = who + where + doing what + how the camera moves.

✗ Bad

"A man runs nervously down the street, very cinematic."

✓ Good

Shot1 side view, slow start; Shot2 knocks over a fruit stall, fast whip-pan to a scared close-up; Shot3 climbs a wall and vanishes, slow pull-back freeze.

Per shot, order it: camera/cut → subject action & expression → position/space change → audio.

💡 Precise timings (e.g. 0-3s) are unstable — don’t hard-limit durations; let pacing follow the story.

③ Action description

Body-part detail + magnitude: hands/legs/head/shoulders with amplitude, speed, force (slowly raise a hand, quickly turn the head).
Prefer slow, gentle, continuous micro-actions; avoid sprinting, big jumps, violent rolls.
Add transitions: state the inertia between actions (raise a hand on the momentum of turning).
Externalize emotion via concrete body detail instead of abstract words like "very sad".

Abstract emotion	Externalized as action & detail
Sadness	Head down, shoulders trembling, red-rimmed eyes, fingers clutching the hem, tears welling but not falling
Joy	Uncontrollable smile, relaxed brows, light steps, humming, spinning in place
Tension / anxiety	Checking the watch, tapping the desk, quick breaths, darting eyes, biting nails
Anger	Clenched fists, tight jaw, heaving chest, sharp glare, words forced through teeth
Relief	A long exhale, shoulders dropping, a faint smile, looking to the horizon

④ Camera language

The model understands standard camera terms well — use them directly: medium shot, close-up, wide shot, slow push-in, steady pan, locked-off shot.

⚠ Use at most ONE camera move per shot; combining push/pull/pan/tilt destabilizes the frame.

⑤ Quality, style & constraints

Quality：HD, rich detail, cinematic, natural color, soft lighting.
Style：Cyberpunk cool blue-purple, retro film, Japanese fresh, 2D anime, 3D guochao.
Constraints：Very important for avoiding artifacts and unwanted elements —

keep it subtitle-free / avoid any text or captions
do not generate a logo
do not generate a watermark

⑥ Special-character conventions

Symbols help the model tell information types apart:

Type	Symbol	Example
Music	（）	(upbeat rock plays in the background)
SFX	<>	<a dog barks in the distance>
Dialogue	{}	{Hello world}; mark the language for non zh/en, e.g. say in Japanese {こんにちは}
Subtitle	【】	[Chapter 1: Departure]

⚠ Keep dialogue in one language; avoid mixing zh/en (proper nouns aside).

⑦ Reference-image strategy

This channel allows up to 2 reference images (imageUrls); camera, pacing and mood are all expressed in text — there is no video / audio input. Spend the limited images on what matters most: a character anchor (lock the look) or a scene setter (lock environment & style).

💡 Typical setups: 1 character image + 1 scene image; or a first-frame + last-frame image (firstFrameUrl / lastFrameUrl). Use a clear single-person portrait, not a multi-view sheet. For sound, set generateAudio to true and describe dialogue / SFX / music with the special-character conventions.

Troubleshooting cheatsheet

Symptom	Fix
Face/ID drift	Provide a face close-up + full-body shot separately, put key assets first; avoid multi-view sheets.
Unwanted subtitles	Add "keep it subtitle-free"; prefer landscape (portrait triggers subtitles more).
Logo / watermark	Add "no logo" and "no watermark".
Style drift	Add explicit style words (e.g. "2D anime"); or convert the reference image to the target style first.
Twins / duplicates	Map each person to a reference (Zhang→Image1); append a global "no duplicate/twin" constraint; use single-person photos; trim the prompt.
Wrong Chinese pronunciation	Replace hard/rare characters with common homophones.
No audio / poor audio	First make sure generateAudio is true; then spell out dialogue {}, SFX <> and music () with the special-character conventions, plus a timbre description. There is no reference-audio input on this channel — timbre can only be described in text.

← Seedance 2 · Introduction Text-to-video API →