/seedance-2-deep-dive

name: seedance-2-deep-dive
description: >
  Deep operating guidance for Seedance 2.0 video generation. Use when selecting
  Seedance 2.0 for a shot, designing multimodal references, writing Seedance-native
  prompts, choosing duration/aspect/quality settings, planning batch generations,
  troubleshooting drift or artifacts, or comparing Seedance 2.0 against Kling, Veo,
  Sora, DoP/Cinema, or other Higgsfield video routes. Complements shot-specifier and
  video-generator by turning Seedance 2.0's multimodal model behaviour into practical
  shot-planning and generation rules.

Seedance 2.0 Deep Dive

Use this skill when `shot-specifier` routes a shot to `seedance_2_0` or when `video-generator` is about to submit Seedance 2.0 jobs through the Higgsfield Model Context Protocol (MCP).

Seedance 2.0 is best treated as a **constraint-driven multimodal video model**, not a text-prompt toy. Text describes the new action. Images, video, and audio references carry identity, style, motion, rhythm, and continuity. The practical skill is deciding which constraints matter, passing them explicitly, and keeping each clip short enough that the model does not drift.

Live-Schema Rule

Before any production generation, inspect the live Higgsfield MCP schema through `video-generator`. Public guidance and creator reports disagree on exact model IDs, quality modes, file limits, duration ranges, and reference roles.

Use the limits below as planning defaults only. If the live MCP schema is narrower, follow the live schema. If the required references cannot be supplied, stop and ask for a production decision.

S01 session 2 observed that the current Higgsfield MCP accepted a Seedance `resolution=1080p` input while still downloading `1344x768` video, and auto-enabled generated audio without exposing a `generate_audio` input key. Treat resolution settings as schema-gated quality hints until the downloaded pixels prove otherwise. Treat audio toggles as intent records unless the live schema exposes them.

When Seedance 2.0 Is The Right Route

Prefer Seedance 2.0 when the shot needs:

• consistent character, costume, product, prop, or recurring visual element identity;
• multiple image references acting as hard creative constraints;
• a short but polished action beat, hook, transformation, product move, or b-roll shot;
• audio-driven pacing from a chosen track, ambience, voice, or sound-effect reference;
• campaign or sequence coherence across many clips;
• image-to-video work from carefully designed start and end frames.

Do not default to Seedance 2.0 when the main requirement is maximum native resolution, long single-shot duration, low-setup one-off generation, exact on-screen text, complex hands, or a large multi-character scene with many competing subjects. Consider Kling for camera-motion-heavy exteriors or motion-control work; consider Veo/native-audio routes when generated audio is the asset rather than an input constraint.

Planning Defaults

| Decision | Default | Reason | |----------|---------|--------| | Duration | 6-8 s first pass; 4-6 s for identity-critical inserts; keep most clips under 10 s | Drift rises with duration; split long ideas into crisp segments | | Upper limit | 15 s only for deliberate hero tests or structured multi-shot prompts | Last seconds are more likely to soften, mutate, or lose continuity | | Aspect ratio | Choose before writing the prompt | Ratio changes composition pressure and what the model emphasizes | | 9:16 | One strong subject, clean background, text safe area if overlays exist | Tall frames push faces and foreground action forward | | 16:9 | Add background control: simple layout, limited background motion, clear negative space | Wide frames invite extra set detail and artifacts | | 1:1 or 4:5 | Product, feed, and commercial detail when supported | Keeps product scale readable without excessive background | | Quality | Draft in fast/medium; final in high only after the shot is coherent | Higher quality sharpens both good detail and bad wobble | | Resolution | Use the manifest's resolution hint for finals when exposed; verify actual pixels after download | S01 current MCP evidence emitted `1344x768` despite a `1080p` Seedance hint | | Batch strategy | Build a shot list and reference plan before spending credits | Random exploration burns budget and weakens continuity |

Multimodal Input Rules

Treat each input as responsible for one job. Avoid overlapping references that ask for different styles, lighting, faces, or motion in the same slot.

Planning limits commonly reported for Seedance 2.0:

• up to 12 files total across images, videos, and audio;
• up to 9 image references;
• up to 3 video references, with short clips preferred;
• up to 3 audio references, with clean, short clips preferred;
• generated output commonly planned in the 4-15 s range.

Verify those values against the live Higgsfield MCP before generation.

Prioritize file slots in this order:

1. Start and end frame anchors when the workflow requires them. 2. Principal character or product identity. 3. Active hero prop and recurring visual elements. 4. Specific location or set layout. 5. Motion or camera reference video. 6. Audio reference for beat, mood, voice, or ambience. 7. Style reference. 8. Supporting detail references.

If the tool exposes input weights, start here:

| Input | Starting weight | Use | |-------|-----------------|-----| | Character/product image | 0.80-0.85 | Exact appearance, costume, object design, brand detail | | Aesthetic/style image | 0.75-0.80 | Colour, lighting, texture, finish | | Environment image | 0.60-0.75 | Location layout and atmosphere | | Motion/camera video | 0.50-0.60 | Camera path, choreography, pacing | | Audio reference | 0.40-0.50 | Mood, tempo, energy, beat timing |

Raise a weight only when that input is underrepresented. Lower it when it dominates the shot or pulls the output away from higher-priority continuity.

Reference Prompting

Us