Seedance prompting

Seedance 2.0 Prompt Engineering

When to load this skill

Load this skill before invoking any media_seedance_* MCP tool or before dispatching to the seedance-director subagent. The default video-router uses this skill's tier-selection guidance to decide which Seedance endpoint to use.

Tier selection

Seedance 2.0 ships exactly two tiers via fal.ai: Fast and Standard.

There is no Pro tier in Seedance 2.0. Any prompt or tool call referencing a Pro tier will fail at the provider level. Do not request it.

Tier	Resolution cap	Best for	Rate (fal.ai)
Fast	720p max (480p also available)	Drafts, A/B variants, iteration cycles, batch volume	~$0.2419/s
Standard	1080p (480p/720p also available)	Production deliverables, hero shots, final renders	~$0.3024/s

Decision rule:

1. Start every new concept with Fast — iterate until prompt + reference set are locked.
2. Promote to Standard only when ready for a production output or when the scene requires 1080p resolution.
3. Native audio is included at both tiers — no surcharge, no separate call.

Cost example: a 10-second 720p clip costs ~$2.42 (Fast) or ~$3.03 (Standard). Always run media_video_cost_estimate before generating to surface actual cost before spending.

Billing note: fal.ai charges per-second of output. For 1080p clips at long durations (Standard tier, ≥10s), a token formula may apply: (height × width × duration × 24) / 1024 × $0.014/1k tokens. fal.ai bills whichever is higher. For 720p clips ≤15s, per-second dominates.

Mode selection — which MCP tool to invoke

Seedance 2.0 exposes three endpoint families (no extend, no lip-sync in v2):

User intent	MCP tool	fal.ai endpoint
Generate video from text only	media_seedance_text_to_video	text-to-video
Animate a start image, optional end frame	media_seedance_image_to_video	image-to-video
Anchor both start AND end frame (frame-morph transition)	media_seedance_image_to_video with endImageUrl	image-to-video
Multi-shot sequence (up to 4 cuts in one clip)	media_seedance_multishot	text-to-video (prompt-structured)
Fuse character/style/motion/audio references	media_seedance_reference_fusion	reference-to-video

extend and lip-sync do not exist as Seedance 2.0 fal.ai endpoints. If a user requests these, surface the gap: for extension, chain multiple t2v/i2v calls; for lip-sync, route to kling-director (Kling V3 Pro has a dedicated lip-sync endpoint).

Prompt structure — 5-part spine (all modes)

Every Seedance prompt benefits from these five ordered layers:

1. Camera — lens height, angle, motion verb. Example: "85mm, eye-level, slow dolly-in". Be specific — Seedance responds to explicit camera verbs, not vague "movement".
2. Scene — location, time-of-day, light source, weather. Example: "rain-slicked Tokyo alley, 2am, sodium vapor pools".
3. Action — subject + verb + object, present tense, no hedging. Example: "woman in white parka crosses a pedestrian bridge".
4. Vibe — mood, film stock, color grade. Example: "contemplative, Kodak Vision3 500T, teal-shifted shadows".
5. Audio — ambient, score, or dialogue cue. Example: "distant train horn, rain on concrete, no music". For reference-to-video with @Audio1, describe what the audio does in the scene.

Compose as a single comma-joined sentence of 30-100 tokens. Do NOT use bullet lists inside the prompt string — the model parses flat prose better.

Worked example (Standard tier, text-to-video)

Spine:

• Camera: "50mm, chest-height, slow truck-left"
• Scene: "neon-lit Hong Kong alley, midnight rain, magenta and cyan signage backlight"
• Action: "delivery courier on bicycle weaves past a fruit stall"
• Vibe: "Wong Kar-wai homage, Fuji 400H push, dreamy halation"
• Audio: "rain on asphalt, distant bicycle bell, no music"

Final prompt:

"50mm chest-height slow truck-left, neon-lit Hong Kong alley at midnight rain with magenta and cyan signage backlight, delivery courier on bicycle weaves past a fruit stall, Wong Kar-wai homage on Fuji 400H push with dreamy halation, rain on asphalt with distant bicycle bell."

Multi-shot syntax (text-to-video, up to 4 cuts, max 15s total)

Multi-shot is achieved via prompt structuring — there is no separate multi-shot endpoint. Use media_seedance_multishot, which builds the structured prompt from a shots[] array.

The timestamp-cut format the model parses:

<global scene direction — style, color, recurring visual anchors, no longer than 80 tokens>

[00:00-00:05] Shot 1: <shot 1 content — camera, action, subject specifics>
[00:05-00:10] Shot 2: <shot 2 content — progression, establish change>
[00:10-00:15] Shot 3: <shot 3 content — payoff or punctuation>

Constraints enforced by media_seedance_multishot:

• Total duration ≤ 15s (Seedance hard limit — any tier).
• Each shot minimum 2s, recommended 3-8s. Shots under 2s produce incoherent motion.
• Up to 4 shots per call. More than 4 produces cut-coherence degradation.
• Global direction ≤ 80 tokens; each shot description ≤ 120 tokens.
• Shots render in array order with zero-frame hard cuts (no crossfade, no dissolve).

Best practice for shot arc:

• Shot 1: wide establishing, slow camera — sets the world.
• Middle shots: subject-focused medium/close — escalate action.
• Final shot: punctuation — hold on detail, pull back to reveal, or hard cut to black.

@-mention reference syntax (reference-to-video)

When using media_seedance_reference_fusion, supply arrays of image, video, and/or audio URLs. fal.ai binds them to placeholder tokens in order:

Array field	Prompt token	Notes
imageUrls[0]	@Image1
imageUrls[1]	@Image2
imageUrls[N-1]	@ImageN
videoUrls[0]	@Video1
videoUrls[N-1]	@VideoN
audioUrls[0]	@Audio1
audioUrls[N-1]	@AudioN

Mention rules:

• ALWAYS @-mention every uploaded reference in the prompt. Unmentioned references are silently ignored by the model.
• Order by narrative importance: identity references first (character face), then style references (environment, material), then motion references (video), then audio last.
• Use natural language — describe what the reference contributes, not just its index: "A woman with the appearance of @Image1, wearing the outfit shown in @Image2, moving with the energy of @Video1, set in the environment from @Image3."

Worked example prompt (3 images, 1 video, 1 audio):

A woman with the appearance of @Image1, wearing the coat from @Image2, moving through
the location shown in @Image3 with the energy from @Video1. Ambient sound from @Audio1.
Slow dolly-in, golden hour side-light, contemplative mood, Kodak Portra 400 grade.

Image-to-video: start frame + end frame transitions

media_seedance_image_to_video accepts:

• imageUrl (required) — the starting frame.
• endImageUrl (optional) — the ending frame.

When endImageUrl is provided, Seedance generates a morph/transition between the two frames. This is the correct way to produce a frame-anchored targeted transition — there is no separate "targeted edit" endpoint in Seedance 2.0. The model interpolates motion, camera, and lighting between the two frames.

Prompt guidance for start→end transitions:

• Describe the motion/transformation you expect between the frames, not the frames themselves (the model sees the images directly).
• Example: "smooth morphing transition, subject turns to face camera, warm interior lighting shifts to cool evening blue".
• Keep prompts under 60 tokens for i2v — the image dominates; the prompt steers motion quality.

Native audio: generation and prompt cues

Both tiers generate audio in the same pass as video (generate_audio: true is the default). The output is a single MP4 with an embedded audio track — no separate ffmpeg merge.

To write prompts that benefit from audio:

Goal	Prompt pattern
Ambient environment	"with the ambient sound of [location]" — e.g. "with ambient forest birdsong"
Cinematic score	"with a slow cinematic strings score" / "with a tense low drone"
Diegetic SFX	Name the sound-producing action: "footsteps on wet gravel", "glass shatters on tile"
Dialogue / voice	For reference-to-video with @Audio1: "voice-over narration from @Audio1 describing the scene"
Silence	"no music, no ambient sound, only the hum of fluorescent lights"

Do NOT include explicit mixing instructions (dB levels, EQ targets) — the model ignores them. Keep audio description to one clause at the end of the prompt.

Director-level camera control vocabulary

Term	Effect
dolly in / dolly out	Smooth depth change — use over "zoom"
tracking shot	Camera follows subject laterally
POV shot	First-person subjective view
rack focus	Shifts focal plane from foreground to background (or reverse)
push-in / pull-out	Aggressive dolly (faster than standard dolly)
pan left / pan right	Horizontal sweep from fixed position
tilt up / tilt down	Vertical reveal from fixed position
crane shot	Rising or falling reveal — wide spatial sweep
handheld	Naturalistic wobble — use for documentary/vérité feel
locked-off	Completely static — composition is the statement
orbit	360 arc around subject — showcase all angles
whip pan	Rapid horizontal sweep — transition energy

Combine at most 2 camera verbs per shot. Three or more collapses into incoherent motion.

Avoid: "zoom" (use "dolly in/out"), "follow" (use "tracking shot"), "move" without direction (always specify axis and direction).

Realistic physics cues

Physics language improves motion coherence:

Domain	Effective cues
Weight	"heavy footfall", "the bag drops with impact", "concrete blocks tumble and slide"
Collision	"billiard balls scatter on impact", "glass fractures from center outward"
Cloth / fabric	"silk dress billows in wind", "leather jacket creases at the elbow", "flags ripple in gust"
Fluid	"liquid pours in a slow arc", "water splashes and beads on surface"
Motion blur	"motion-blurred background, sharp subject", "frozen-moment freeze with trailing blur"

Add one physics clause per shot for any scene with significant physical interaction. Omitting physics cues for contact scenes leads to floaty/weightless motion.

Resolution and duration reference

Parameter	Valid values	Notes
resolution	"480p", "720p", "1080p"	1080p is Standard tier only
duration	"auto", "4", "5", ..., "15"	String enum, max 15s
aspectRatio	"auto", "21:9", "16:9", "4:3", "1:1", "3:4", "9:16"	Default "auto"

Default: resolution="720p", duration="auto", aspectRatio="auto", generateAudio=true.

Use "9:16" for vertical social (Reels, TikTok, Shorts). Use "21:9" for cinematic widescreen. Use "1:1" for square format social.

Common failure modes

Symptom	Cause	Fix
References silently ignored	Missing @-mention in prompt	@-mention every uploaded ref by its token (@Image1, @Video1, etc.)
Multi-shot shots blend together	Timestamp markers malformed	Use exact [MM:SS-MM:SS] Shot N: format; no variation
Motion crawls / floaty	No physics cues, generic verbs	Add specific camera verb + physics clause for contact scenes
Audio mismatched to visuals	Audio description placed mid-prompt	Move audio cue to the last clause in the prompt
Face drifts across multi-shot	No persistent identity ref	Use reference-to-video with @Image1 for character identity across shots
i2v output ignores end frame	endImageUrl not semantically described in prompt	Add a transition description clause: "morphing into", "shifting to"
Cost overrun at 1080p long clip	Token billing kicks in	Recompute: (1080×1920×duration×24)/1024 × $0.014 and compare to rate × duration

When NOT to use Seedance

• User needs lip-sync to audio → route to kling-director (Kling V3 Pro, dedicated lip-sync endpoint)
• User needs video extension beyond 15s → chain multiple Seedance calls or route to kling-director (video-extend)
• User wants Higgsfield Soul ID persistent character identity → route to higgsfield-director
• User needs sub-second or real-time preview → Seedance is async queue-based; use a synchronous provider

Legal note — IP risk

Seedance 2.0 is classified ipRiskLevel: high in the media-forge model registry (Disney/Paramount C&D context). The operator running media-forge assumes full responsibility for compliance with applicable IP law. Do NOT include recognizable IP-protected characters by name or likeness in prompts (e.g. Disney characters, Marvel heroes, Paramount franchises). Do NOT generate content featuring real persons without explicit consent context. This is the only surface in media-forge that mentions IP risk — there is no runtime gate.