Applies to: Users on the Seedance video model. The controls below are specific to Seedance and are not available on every Venice video model.
Seedance adds multimodal controls that go beyond a plain text prompt: you can assign reference assets specific jobs, choose how much references control the result, edit with first/last frames, extend clips, and score audio with a dedicated syntax.
Assign reference authority (@-assets)
Each reference you upload gets one explicit job. Number them in upload order:
@Image1 defines identity, prop, location, or style.
@Video1 defines only motion, camera, timing, or cuts.
@Audio1 defines only voice, dialogue, music, ambience, or effects.
Rules:
One asset = one explicit job. Don't ask one image to define both identity and style.
Reject what you don't want carried over (unwanted people, clothing, setting, text, style, audio).
If several views show one subject, say so — and lock the exact count.
When sources conflict, name priority by attribute.
Choose the workflow (locked vs. unlocked)
Unlocked — references guide the result (subject, motion, style, audio). Use Storyboard or Keyframes.
Locked — inputs control the timeline or geometry. Use Edit, First/Last frame, or Extension.
Storyboard = loose plot plus shot order (about 15 clean panels or fewer).
Keyframes = closer visual alignment; ordered independent images.
Camera and action
One dominant camera move per shot. Say "slow dolly in at eye level," not "zoom in cinematically."
Chain the cause and effect: trigger → contact/evasion → reaction → recovery → next decision.
Audio syntax
( music ) < effects / ambience > { dialogue } [ subtitles ]
Assign speaker, language/accent, and delivery.
Explicitly reject unwanted reference audio, subtitles, or BGM.
30-second story map (for longer single generations)
0–6s establish subject, location, situation, direction
6–12s trigger changes the situation
12–18s reaction, transformation, or maneuver
18–25s one coherent cause-and-effect exchange
25–30s resolution, aftermath, readable final frame
Rules: one primary state change plus one visible end state per stage; consecutive, non-overlapping integer-second ranges; timestamps allocate time, they are not frame-accurate cuts.
Continuity locks
Track exact count, identity, clothing, prop owner and hand, position, screen direction, environment, lighting/weather, damage/state, camera axis, and audio source. For handoffs, state BEFORE → transfer moment → AFTER. Make every new stage begin from the previous stage's visible end state.
Copy-ready template
@Image1 defines <subject identity/appearance>.
@Video1 defines only <motion/camera/timing>; do not inherit <people, setting, text, style, audio>.
@Audio1 defines only <voice/music/ambience/effects>.
<Subject> performs <one visible event> in <environment>.
0-<time>: <first change and visible end state>.
<time>-<time>: Continue from that state; <second change and final state>.
Camera: <one dominant direction per shot>.
Audio: <assigned dialogue, music, effects, and ambience>.
Keep <identity, count, clothing, ownership, layout, lighting, and direction> consistent.
No <duplicates, extra people, text, logos, unwanted audio, or unexplained cuts>.
Seedance 2.5 specs: up to 30s video, 50 total references, 30 images (up to 4K), 10 videos / 30s combined, 10 audio / 30s combined, ratio 0.4–2.5, 10+ languages.