Multimodal Input Tips: Make AI Understand Your "Director Intent"

Multimodal Is Not Simple Stacking

Seedance 2.0 multimodal input is not “more assets equals better.” What matters is giving each modality a clear role. A strong multimodal prompt is like a concise storyboard.

Four-Modality Division Model

Modality	Best Use	Recommended Count
Text	Narrative logic, detail, emotional tone	1 concise instruction
Image	Visual style, character look, scene composition	3–5 images
Video	Camera movement, action rhythm, transition style	1–2 clips
Audio	Music style, sound character, speech rhythm	1 clip

Scenario 1: Brand Advertising

Goal: Generate a 15-second product showcase video

• Images: 3 high-res product shots (front, side, in-use scene)
• Video: Reference pacing from an existing brand ad
• Audio: Reference brand ad soundtrack
• Text: “Minimal premium style — product slowly lights up from darkness, 360° rotation, final hold on logo”

Scenario 2: Narrative Short

Goal: Generate a 12-second story-driven clip

• Images: 2 character sheets + 2 scene concept images
• Video: Reference cinematic camera work from a film clip
• Text: Full scene description with setup, development, and payoff

Scenario 3: Music MV Segment

Goal: Audio-visual sync for dance/performance

• Audio: Target song segment (model reads rhythm)
• Video: Reference dance moves or performance style
• Images: Stage/scene visual references
• Text: Supplement lighting and atmosphere

”Director Mindset” for Prompts

1. Set the tone first: What emotion does this video carry?
2. Define structure: How many shots? How long each?
3. Then details: Specific content per shot
4. Finally style: Visual and audio character

Debugging Tips

If output drifts from expectation, troubleshoot in this order:

1. Check for contradictions between modalities
2. Reduce reference count and add back one by one
3. Split text instructions into more specific descriptions
4. Use video editing to fine-tune instead of full regeneration