Multimodal control
Combine text, images, video, and audio in one generation. Use each file for a specific role: character, environment, camera rhythm, motion, audio, or style.
AI video workflow
Create pro AI video with Seedance 2.0.
Seedance 2.0 is a multimodal audio-video model built for structured directing. Use text, image, video, and audio references together, then guide the model with clear roles, physical action, and shot-based timing.
9image references
3video references
3audio references
Mental model
Treat Seedance like a controlled directing system, not a poetic toy. Every reference needs a job, and every shot needs one readable action.
What Seedance 2.0 does best
It shines when you stop writing vague visual descriptions and start directing the model with scene logic, reference roles, physical motion, and clean constraints.
Shot continuity
Build scenes with a beginning, development, and payoff. A short beat structure gives the model temporal logic and keeps motion readable.
Physical action
Describe what bodies, objects, light, dust, fabric, vehicles, and environments actually do on screen. Tangible behavior beats abstract mood.
Core operating principles
These are the rules that keep Seedance generations controlled, readable, and cinematic.
One shot, one dominant action
Focus each frame on a single clear narrative or action point. Visual clutter lowers impact and makes the model guess.
Camera is not decoration
The camera path should serve the subject, story, or emotion. Do not add camera moves just because they sound cinematic.
Action must be physical
Prioritize visible behavior: falls, collisions, dust, fabric, fire, hands, eye movement, product interaction, and body mechanics.
Environment is active
The location should shape the action. Treat furniture, light, weather, crowds, streets, glass, walls, and props as part of the scene.
References need roles
Define exactly how external information is used. A reference can control identity, wardrobe, palette, composition, motion, rhythm, or ambience.
Structure beats length
A shorter prompt with hierarchy usually beats a long prompt full of competing ideas. Organize first, expand only where needed.
The Seedance workflow
Use this repeatable process before every generation, no matter which platform gives you access to Seedance 2.0.
Step 01
Choose the format
Decide whether you need a single-shot scene, multi-shot story, reference-based generation, continuation, POV, edit, extension, or detail shot.
Step 02
Upload references with roles
Never upload files as decoration. Label what each reference should control: identity, wardrobe, palette, composition, camera motion, blocking, speech rhythm, or ambience.
Step 03
Write one dominant action
Each shot should focus on one clear event. Avoid stacking too many actions, camera moves, style words, and character changes into the same moment.
Step 04
Build the shot script
For advanced results, write timecoded beats. Use a structure like establish, develop, payoff so the model understands progression across time.
Step 05
Generate, review, refine
If the result is chaotic, simplify the prompt. If style is ignored, adjust style weight. If identity drifts, fix the reference role. If camera resets, lock the camera path.
Main prompt modes
Choose the mode before writing. The mode determines how much control you need and how you should structure references.
Single-shot
One continuous readable event. Best for live performances, interviews, simple actions, and unedited footage.
Multi-shot
Linked beats across time. Best for transformations, mini narratives, product launches, explainers, and event highlights.
Reference-based
Use a source image, video, or audio file to guide character, scene, motion, camera, rhythm, or style.
Continuation
Extend a story or scene from a previous output while preserving continuity and avoiding repeated action.
Edit
Modify specific elements of an existing image or clip while keeping the rest of the scene stable.
Extension
Expand the canvas or scene from an image while preserving the original visual logic.
POV
Change the visual perspective or camera angle so the viewer experiences the scene through the subject.
Insert / detail shot
Add a detailed close-up or element shot to make the scene feel more complete and cinematic.
Reference roles
Seedance 2.0 can read multiple reference types, but every reference should have one job.
Image reference
Use images for identity, wardrobe, palette, composition, environment, product design, pose, or first/last frame anchors.
Video reference
Use video for motion rhythm, camera behavior, blocking, pacing, transition feel, choreography, or action timing.
Audio reference
Use audio for timing, speech rhythm, ambience, music energy, impact cues, or the emotional rhythm of the scene.
Bad: Use @Video1 for the scene.
Good: Reference @Video1 for camera movement only. Use @Image1 for character identity and @Image2 for environment.
Prompt formula
Use this as the fast version when you do not need a full shot script.
[Subject], [Action], in [Environment], camera [Camera Movement], style [Style], avoid [Constraints]
Subject
Who or what appears. Be specific about visual features, wardrobe, identity, and role.
Action
What happens physically. Use precise verbs and visible motion instead of abstract emotion.
Environment
Where it happens. Include light, atmosphere, space, and objects that can interact with the action.
Camera
Use one primary movement: push-in, tracking, orbit, handheld, aerial, fixed, pan, or pull-out.
Style
Anchor the visual feel without creating adjective soup. Keep it focused and consistent.
Constraints
Tell the model what to avoid: jitter, bent limbs, identity drift, flicker, chaotic composition.
Multi-shot beat structure
For stronger narrative clips, use a simple three-beat structure instead of a loose paragraph.
Beat 1
Establish
Show the subject, environment, relationship, or problem. The viewer should understand the scene immediately.
Beat 2
Develop
Introduce motion, conflict, transformation, product use, camera progression, or a visible change in the scene.
Beat 3
Payoff
End with the result: impact, reveal, product moment, final pose, explosion, joke, reaction, or clean cinematic finish.
Continuation logic
When extending a clip, the model needs anchors. Tell it what to preserve and what must happen next.
Use a last-frame anchor
The continuation should start immediately after the final frame, not replay the previous action.
Use identity anchors
Preserve character, outfit, environment, emotional state, and camera direction unless you intentionally change them.
Use previous clip as memory
The previous video should guide continuity, not trap the model into repeating the same beat.
State what must not repeat
Add direct constraints like “do not repeat the previous action” or “continue from the final pose.”
Common mistakes
Most bad generations come from too much ambiguity, not from a weak model.
Prompt overload
Too many actions, references, camera ideas, and style words compete with each other. Remove everything that does not serve the main shot.
References with no roles
Do not just upload files and hope the model understands. Say exactly what each reference controls.
Camera as decoration
Camera movement must serve the story. Use one primary camera idea and keep subject movement separate from camera movement.
No hierarchy
Make the prompt easy to parse: style, duration, shot beats, consistency rules, and negative constraints.
Adjective soup
Too many style words dilute the result. Use one strong style anchor, then describe physical details that support it.
Contradictory identity
If character, outfit, age, face, or role references conflict, Seedance may drift. Give one clear identity source.
Changing too many variables
When refining, change one thing at a time: camera, style, reference, seed, duration, or prompt structure.
Literary emotion instead of behavior
“She feels powerful” is weaker than “she stands still, lifts her eyes, tightens her grip, and steps forward.”
Troubleshooting by symptom
Use this when a generation looks close but not quite usable.
Chaotic result→Refine prompts and seeds
Style ignored→Adjust style weights and simplify style language
Character drift→Re-seed or fix identity reference roles
Camera resets→Lock the camera path and reduce competing camera moves
Previous beat replays→Trim or edit input, then state the next action clearly
Motion feels fake→Increase realism settings and describe physical mechanics
Clip feels unstable→Stabilize footage or tracking references