seedance-prompting

name: seedance-prompting description: Prompt + parameter playbook for ByteDance Seedance 2.0 image-to-video on fal.ai. Use whenever generating, regenerating, or rewriting prompts for VID-shot assets where the provider is seedance-fal-img2vid. Pairs with webapp/lib/agents/providers/fal-seedance.ts and webapp/components/vgen/ProviderControlPanel.tsx. status: ACTIVE owner: EXEC-VGEN (Animator) applies_when: agent: [EXEC-VGEN] hard: false maturity: stub-v0.1 created: 2026-05-14

Seedance 2.0 — Prompting Playbook (SandyStudio)

v0.1 — Sprint β kickoff. Will grow with each E2x production. Append findings as labelled subsections; do not rewrite history.

When this skill applies

• The shot's resolved provider is seedance-fal-img2vid (default for character_video per migration 0028).
• The agent (EXEC-VGEN or the regenerate-video endpoint) is composing a prompt for the queue submission.
• Director is hand-editing the prompt textarea in <VGENShotPanel> and wants guidance.

For Veo 3 the rules are different — see future veo-prompting skill. Do not cross-pollinate.

The seven prompt slots

Seedance 2.0 responds best to a structured prompt, not a paragraph. The following slot order produces the most consistent character motion and camera control in our internal probes (E20 / 2026-05-13):

SHOT TYPE:
<framing label — e.g. "Medium close-up, 16:9">

SUBJECT:
<who/what is visible, anchored to the locked reference image>

ACTION:
<one clear physical action; verbs only — Seedance over-interprets adverbs>

CAMERA:
<lens framing + movement; static unless motion is the point>

MOTION:
<character motion + object motion + environmental motion, separated>

STYLE:
<visual style + lighting + rendering — match Series Bible style canon>

CONTINUITY:
<locked character / location / costume / props — refer back to anchor>

NEGATIVE:
<what must NOT happen — comma-separated list>

Slot omission is allowed but prefer all seven when the shot is final quality. Skipping CONTINUITY is the single most common cause of character-drift defects we've seen.

Worked example (E20 SS-S14-E20-A1-SC03-SH01)

SHOT TYPE:
Medium close-up, 16:9, eye-level framing.

SUBJECT:
Sandy (animated hourglass character), tilting forward over a pale cream-white
counter inside a softly-lit perfume boutique.

ACTION:
Sandy leans forward. Both eyes water. One fat cartoon tear forms at the
inner corner of each eye. He smiles wider and leans in further.

CAMERA:
Static camera, slow 5% push-in. Keep Sandy centred. No handheld shake.

MOTION:
Subtle character animation only: eyebrow lift, tear bead descent,
shoulder lean. Mist particles drift unprompted off Perfume Madame in the
right edge of frame. Background remains stable.

STYLE:
Stylised 2D animation, S14 STYLE CANON v1.1 (outline-only pencil edge,
flat vector fills, no hatching). Warm cinematic lighting.

CONTINUITY:
Match reference image exactly for Sandy's proportions, golden sand colour,
dark-grey cap, and the perfume counter layout. Perfume Madame's atomiser
remains gold-chain. No outfit changes.

NEGATIVE:
no extra limbs, no face morphing, no costume changes, no extra characters,
no text or logos, no camera shake, no melting objects, no hatching.

Parameter playbook

Cost intuition (Sprint β baseline)

Plan accordingly when iterating. Standard 1080p × 15s = ~$10. Use 480p for camera tests; switch to 1080p when locked.

Hard rules (don't violate)

1. Always pass generate_audio: false — Seedance inline audio clashes with the music track laid down by EXEC-MGEN (Composer) and the muxed final cut from EXEC-STITCH (Online Editor). Already enforced in fal-seedance.ts line ~195; surfaced here so prompt authors never ask for audio cues in the prompt.

2. Never describe a character's identity in the prompt as if it were novel. The character lives in the locked reference image. The prompt describes what they do, not who they are. Example:

   • ❌ "Sandy, a sand-filled hourglass character with golden granules…"
   • ✅ "Sandy leans forward; golden sand surges warmly upward into his chest."

3. Mention the reference asset's framing. If the reference is a wide shot but the prompt asks for a close-up, Seedance interpolates an ungrounded zoom — character proportions drift. Match prompt framing to the reference, or supply an end_image_url that conveys the close-up target.

4. One primary causal chain per shot, not one verb (softened 2026-05-27). Beats may chain causally («tap → launch → smash → vibrate» — four beats, one subject, one chain) and a reactive micro-beat on a secondary subject is fine (e.g. character recoil from the primary impact). Seedance 2 renders these without blur. What DOES blur Seedance is genuinely parallel independent actions — split those into two shots. Comedy rhythm (technology.md §3.5) remains 3-5s per cut.

5. Negative prompt is not optional. At minimum: no extra limbs, no face morphing, no costume changes, no extra characters, no melting objects, no text, no logos. Seedance is robust but the surcharge of including this list is zero and it visibly reduces defect rate (E20 probe 2026-05-13).

6. Reference/start-image ASPECT must match the render aspect_ratio (2026-06-09, E03 Shorts probe). A 16:9 landscape ref (1536×1024) into a 9:16 render makes Seedance crop ~44% horizontally and recompose the frame → content loss + identity drift. This is distinct from rule 3 (framing): rule 3 is wide-vs-close, this is landscape-vs-portrait. For vertical delivery (Shorts / Reels / TikTok) the reference MUST be authored portrait (1024×1792 per the EREF Designer size table). If you receive a landscape ref for a 9:16 shot, STOP and flag upstream — do NOT render. Root cause on E03: the episode had no delivery_targets, so the EREF Designer silently fell back to youtube_landscape and authored landscape refs.

Known quirks (collected from production probes)

• Parent-truncated status URLs — fal returns …/seedance-2.0/requests/<id>/status even though POST was to …/seedance-2.0/image-to-video. Use returned URLs verbatim; do not reconstruct. Already handled in adapter.
• Duration must be string enum, not int — payload uses "5", not 5. Adapter handles. Don't fight this.
• Seedance enforces 4–15s clamp. Storyboard timings outside that range are silently clamped at submit time; runner logs the clamp.

Open questions (to refine with more E21+ probes)

• Best resolution for the regenerate-video iteration loop — currently we default to 720p but a Director ablation on 1080p iteration cost-vs-quality is pending.
• Whether seed-locking across a scene (same seed, ±10s of prompt drift) consistently keeps character identity. Anecdotal yes, no formal probe.
• end_image_url for emotion arc (frame 0 = neutral, frame N = punchline) — promising but unconfirmed for our gag density.

Cross-references

• Adapter: webapp/lib/agents/providers/fal-seedance.ts
• Capability manifest: webapp/lib/api/provider-capabilities.ts
• UI: webapp/components/vgen/ProviderControlPanel.tsx
• Phase A.1 prompt builder: webapp/lib/api/vgen-shot-helpers.ts (currently provider-agnostic; future fork into seedance-specific composer is a follow-up)
• Series style anchor: technology.md §3.5 (shot rhythm)