Structured Captioning

导演模式 · 结构化描述(HoloCine)

Overview

把提示词拆成 GLOBAL(场景/时间/天气/情绪)→ CHARACTER DEFINITIONS([char1]/[char2] 锚点)→ PER-SHOT(只用标签引用,不重描外貌),省 ~40% token 并锁定唯一外观。

When to Use

• 用:多格反复重描角色外貌、token 浪费、外观出现微漂移。
• 不用:单图单角色无复用;锚点本身的提取质量(→ director-anchor-extraction-quality)。

STRUCTURED CAPTIONING — separate global context from per-shot details (HoloCine pattern):

Prompt Structure (strict order):

1. GLOBAL: scene environment, time of day, weather, overall mood
2. CHARACTER DEFINITIONS: [char1]: full appearance anchor, [char2]: full appearance anchor
3. PER-SHOT: each panel references characters by tag [char1], never repeats full appearance

Why This Matters:

• Repeating character appearance in every panel prompt causes micro-variations → inconsistency
• Tag references ([char1]) force the model to maintain one canonical appearance
• Reduces total token count by ~40%, leaving more capacity for scene details

Per-Shot Format: Panel N: [shot type], [char1] does X while [char2] does Y, [lighting], [composition]

Anti-Patterns:

• DO NOT repeat "a young woman with long black hair wearing a red coat" in every panel
• DO NOT mix character definitions into per-shot descriptions
• DO NOT omit the global section — it anchors the entire sequence

Example

GLOBAL: dim cyberpunk bar, night, rain outside, smoky neon mood. CHARACTERS: [char1]: 30s woman, silver bob, scar on left cheek, black trench. [char2]: old bartender, bald, grey beard, maroon vest.

• Panel 1: wide, [char1] enters, [char2] polishes a glass, teal rim light.
• Panel 2: close-up, [char1] slides a coin, [char2] raises an eyebrow. → Appearance is never re-described per panel — only [char1] / [char2] tags carry it, saving tokens and locking one look.

Common Mistakes