Prompt Engineering for Video

The biggest mental shift required for effective Seedance prompting is moving from "describing an image" to "directing a scene." Amateur prompts describe what things look like. Professional prompts specify what the camera does, where the light comes from, how subjects move, and what emotion the shot should evoke.

The directorial prompt formula: [Subject] + [Action/Motion] + [Camera Behavior] + [Environment] + [Lighting/Atmosphere] + [Style/Aesthetic]

Example - weak prompt: "A chef cooking in a kitchen"

Example - directorial prompt: "A weathered chef in a white apron carefully plates a dish, close-up shot slowly pulling back to reveal the bustling kitchen behind, warm tungsten lighting from overhead, cinematic depth of field, golden hour color grade, 35mm film aesthetic"

The second prompt gives Seedance far more constraints to work within, which paradoxically produces more interesting and controllable results. The model has less ambiguity to fill in randomly.

Shot type vocabulary to use in prompts:

• Extreme close-up (ECU) - eyes, hands, small objects
• Close-up (CU) - face and shoulders
• Medium shot (MS) - waist up
• Full shot (FS) - entire body
• Wide shot / establishing shot - subject in environment
• Aerial/drone shot - top-down perspective
• Dutch angle - tilted camera for tension

Lighting terminology that Seedance responds to well: golden hour, blue hour, overcast diffused, harsh directional, rim lighting, chiaroscuro, neon-lit, practical-only lighting, day interior soft box, night exterior practical.

Camera movement is one of the highest-leverage elements in AI video prompting. Seedance supports natural language camera direction, and using the right terms unlocks dramatically more cinematic outputs.

Supported camera movements in Seedance (use these exact phrases):

Lateral movements:

• "camera pans left/right" - horizontal pivot from a fixed point
• "camera tracks left/right" - parallel lateral movement alongside subject
• "camera dollies forward/backward" - physical push toward or away from subject

Vertical movements:

• "camera tilts up/down" - vertical pivot from a fixed point
• "camera cranes up/down" - physically rising or lowering (heroic/ominous effect)
• "aerial shot descending slowly" - drone-style descent

Focal movements:

• "slow zoom in" - optical zoom creating intimacy
• "slow zoom out" - reveal shot, building context
• "rack focus from foreground to background" - selective focus shift

Dynamic movements:

• "handheld shaky camera" - documentary/guerrilla feel
• "smooth steadicam follow shot" - tracking subject from behind
• "360-degree orbit around subject" - surround/hero shot
• "whip pan" - rapid lateral pan for transitions

Combining movements: The most cinematic results come from combining motion types. "Camera slowly pushes in while tilting up slightly, subject remains centered" describes a compound move that creates a rising, intimate feel.

The surround shot hack: Seedance handles "camera slowly orbits clockwise around the subject" very well for product shots, character reveals, and architectural showcases. This is one of the most reliably impressive single-prompt moves you can specify.

Seedance 2.0 introduced a major capability: multi-shot generation in a single prompt. By using the phrase "lens switch" in your prompt, you instruct the model to create a visual cut to a new shot - while maintaining subject identity, scene consistency, and stylistic continuity.

The lens switch syntax: [Shot 1 description]. Lens switch. [Shot 2 description]. Lens switch. [Shot 3 description].

Example: "A woman in a white dress runs through a wheat field at golden hour, wide establishing shot, camera tracking alongside. Lens switch. Close-up of her face, wind in her hair, joyful expression, shallow depth of field. Lens switch. Aerial shot pulling back to reveal the entire field at sunset, camera craning upward."

This generates a three-shot sequence with cinematic coherence - something that previously required generating and stitching three separate clips.

Rules for effective lens switches:

1. Keep subject description consistent across shots (same clothing, hair, name if using a reference)
2. Specify the shot type explicitly after each lens switch
3. Don't change location dramatically between shots - the model maintains scene context
4. Limit to 3 shots maximum for 8-second clips; 2 shots for 4-second clips
5. Maintain consistent style descriptors ("cinematic," "film grain," "color grade") throughout

Scene continuity tips: If you're generating a series of clips that should feel like they belong to the same film, include a consistent "style signature" in every prompt - for example, "35mm film grain, desaturated teal-orange color grade, overcast natural light" - and all your clips will feel visually unified even if generated separately.

Professional AI video workflows are never single-shot. The best creators use a structured iterative process that systematically narrows in on the desired result while minimizing wasted credits.

The 4-stage iteration framework:

Stage 1 - Concept validation (Lite tier): Write a minimal prompt with just subject, action, and environment. Generate 2-3 variations at low cost. This validates the core idea before investing in expensive generations.

Stage 2 - Style locking: Take the most promising Stage 1 result and add detailed style, lighting, and camera descriptors. Generate 2 variations. Identify which style elements are working.

Stage 3 - Motion refinement: Focus exclusively on camera movements and subject motion. Adjust timing words ("slowly," "quickly," "gradually") and camera behavior. Generate 1-2 variations.

Stage 4 - Final polish (Pro or 2.0 tier): Combine all winning elements from stages 1-3 into a single comprehensive prompt. Generate at full quality.

Prompt element prioritization: When outputs aren't matching your intent, identify which element to change first. Common failure modes and their fixes:

• Wrong style/mood - adjust lighting and color grade terms
• Jerky or unrealistic motion - add "smooth," "natural," "fluid" to motion descriptors
• Wrong composition - specify shot type explicitly (close-up, wide shot, etc.)
• Subject inconsistency - use a reference image in I2V mode instead of T2V

Seed/variation control: Most Seedance implementations let you fix the generation seed to reproduce similar outputs while varying the prompt. Use this to isolate the effect of single prompt changes.