← 返回 Skills 市场
beermanzz

Character Consistency Workflow

作者 beermanzz · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ✓ 安全检测通过
40
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install character-consistency-workflow
功能描述
AI 短剧/短片创作中的角色一致性工作流。解决"人物漂移"(character drift)问题——多镜头生成时同一角色外观漂移、换脸、换装。触发关键词:角色一致性、人物漂移、换脸、定妆照、AI视频、多镜头、一致性Tokens、Seedance、Nano Banana。当需要保持角色跨镜头外观统一、防止AI生成时...
使用说明 (SKILL.md)

Character Consistency Workflow

AI 短剧创作 · 角色一致性工作流

解决 AI 生成多镜头时"人物漂移"(Character Drift)的系统方法论。


🔍 问题诊断:什么是 Character Drift?

Character Drift(人物漂移) 是 AI 生成视频序列时,同一角色在不同镜头间的外观变化:

漂移类型 表现 严重程度
长相漂移 脸型、五官、发色在镜头间变化 ⭐⭐⭐ 高
服装漂移 衣服颜色、款式、饰品不一致 ⭐⭐ 中
表情漂移 情绪表情、眼神光变化 ⭐⭐⭐ 高
肤色漂移 肤色明暗、冷暖色调跳变 ⭐⭐ 中
年龄感漂移 面部年龄特征不稳定 ⭐⭐ 中

三大一致性层级

Level 1: Appearance(长相一致性)
  └── 脸型、五官比例、肤色、发型的基准锁定
Level 2: Costume(服装一致性)
  └── 服装颜色、款式、饰品、道具的连续性
Level 3: Emotion(表情一致性)
  └── 情绪状态、眼神、面部肌肉走向

技术背景

  • Seedance 2.0 三大核心突破之一:解决 Character Drift(漂移角色)
  • Nano Banana 2:支持 5 角色 / 14 物体一致性
  • 当前主流工具(LibLib、即梦等)均存在不同程度的漂移问题,需要系统性防漂移工作流

📦 第一阶段:Asset Preparation(资产准备)

定妆照拍摄规范

每个角色需要准备以下参考图(详见 references/asset-preparation.md):

  1. 正面定妆照 × 1

    • 纯色背景,正面机位,自然光
    • 完整上半身可见
  2. 侧面/3/4 侧面照 × 2

    • 与正面照同天拍摄,光源一致
    • 服装、妆容完全相同
  3. 表情参考 Sheet × 1 组

    • 6-8 种基础情绪表情
    • 平静、开心、愤怒、悲伤、惊讶、沉思
  4. 服装细节特写 × 3-5 张

    • 领口、袖口、饰品、面料纹理
    • 用于补全镜头细节

文件夹组织

characters/
├── protagonist/
│   ├── ref_front.jpg      # 正面定妆照
│   ├── ref_side.jpg       # 侧面照
│   ├── ref_3quarter.jpg   # 3/4侧面照
│   ├── expressions/
│   │   ├── neutral.jpg
│   │   ├── happy.jpg
│   │   ├── angry.jpg
│   │   └── ...
│   └── costume_detail/
│       ├── collar.jpg
│       └── accessory.jpg
└── antagonist/
    └── ...

🏷️ 第二阶段:Consistency Tokens(一致性 Tokens)

Token System 概述

Consistency Tokens 是在每个 prompt 中重复使用的固定描述块,用于锁定角色特征。

Token 公式:

[BASE_IDENTITY] + [PHYSICAL] + [CLOTHING] + [HAIR] + [DISTINGUISHING_MARKS]

完整 Token 模板

Token: CHARACTER_A

BASE_IDENTITY:
- 姓名/角色名:[角色名]
- 年龄感:青年/中年/少年
- 性别外观:[描述]

PHYSICAL:
- 脸型:[如:鹅蛋脸,颧骨略高]
- 眉型:[如:剑眉,眉峰分明]
- 眼型:[如:杏眼,眼尾微翘]
- 鼻型:[如:直鼻,鼻梁挺拔]
- 唇型:[如:薄唇,唇角微扬]

SKIN:
- 肤色:[如:健康小麦色,质地细腻]
- 特殊:[如:左眼角一颗泪痣]

HAIR:
- 发长:[如:长发及腰]
- 发色:[如:深棕色]
- 发型:[如:中分,发尾微卷]

CLOTHING(按场景固定):
- 场景A服装:[精确描述,包括颜色、款式、饰品]
- 场景B服装:[精确描述,包括颜色、款式、饰品]

EXPRESSION_PATTERN:
- 常用表情:[如:平静时眉心微蹙]
- 眼神特点:[如:目光沉稳,偶有闪躲]

多场景 Token 示例

# === SCENE 1 TOKENS ===
[Character: LINDA, female, 28, oval face, sharp chin, arched eyebrows, 
almond eyes, high nose bridge, thin lips, fair skin with rosy cheeks, 
long black hair in low ponytail, wearing navy blue blazer, white silk 
blouse, pearl earrings, minimal gold watch on left wrist]

# === SCENE 2 TOKENS (same character, different outfit) ===
[Character: LINDA, female, 28, oval face, sharp chin, arched eyebrows, 
almond eyes, high nose bridge, thin lips, fair skin with rosy cheeks, 
long black hair in low ponytail, wearing red cocktail dress, black stiletto 
heels, diamond drop earrings]

注意:长相描述(Physical)不变,只有服装描述(Clothing)随场景更新。


🎬 第三阶段:Cross-Shot Consistency(跨镜头一致性)

Prompt 构建规则

规则 1:每个 Prompt 包含完整 Token

即使镜头只拍背影,Token 也必须完整存在,防止特征漂移。

规则 2:Reference Image 路径写入 Prompt

Style Reference: ./characters/protagonist/ref_front.jpg
Character Identity locked: [Token Block]

规则 3:Negative Token(反向描述)

Negative: [避免的特征,如:different face shape, different hair color, 
different clothing, extra limbs, deformed hands, blurry face]

规则 4:镜头间的渐变规则

镜头切换类型 Token 处理
同一场景内切 完全一致的 Token
场景跳切 服装更新,Physical Token 不变
时间跳跃 重新确认所有 Token,生成前对照定妆照
情绪特写 在 Token 基础上叠加表情描述(emotion overlay)

表情情绪叠加格式

# Base Token + Emotion Overlay
[Character: LINDA ...] + Emotion: eyes wide, pupils dilated, 
mouth slightly open, eyebrows raised, expression of shock

镜头序列示例

Shot 1: Medium Shot
[Character: LINDA ... wearing navy blazer...] + neutral expression
Prompt: Medium shot, LINDA standing in office, facing camera, 
natural lighting, cinematic

Shot 2: Close-up (same scene)
[Character: LINDA ... wearing navy blazer...] + slight frown, focused eyes
Prompt: Close-up, LINDA face, eyes looking at document, slight frown, 
cinematic lighting, shallow depth of field

Shot 3: Cutaway to hands
[Character: LINDA ... wearing navy blazer...] + (no face needed)
Prompt: Close-up, LINDA hands typing on keyboard, navy blazer sleeve visible, 
gold watch on wrist, cinematic lighting

✅ 第四阶段:Quality Check(质量检查)

一致性检查清单

生成完成后,按以下顺序检查:

□ 1. 长相一致性
  □ 脸型是否与定妆照一致?
  □ 眉毛形状、眼睛形状是否漂移?
  □ 发色、发长是否一致?

□ 2. 服装一致性
  □ 服装颜色是否一致?
  □ 饰品位置是否正确?
  □ 面料质感是否一致?

□ 3. 肤色一致性
  □ 肤色明暗是否跳变?
  □ 冷暖色调是否一致?

□ 4. 比例一致性
  □ 头身比是否稳定?
  □ 手部比例是否正常?

□ 5. 光影一致性
  □ 光源方向是否一致?
  □ 光影质感是否统一?

漂移问题修复对照表

问题 修复方案
脸型漂移 在 Token 中加入更精确的脸型描述,使用 Reference Image
发色漂移 明确标注发色 hex 值,如:深棕色 #3B2417
服装颜色漂移 明确标注服装颜色,如:海军蓝 #1B2838
眼睛形状漂移 使用 expression sheet 中的眼部特写作为参考
手部畸形 加入 negative token: deformed hands, extra fingers
饰品消失 在 Token 中详细列出所有饰品,加入场景描述

跨平台一致性策略

平台/模型 一致性能力 建议
Seedance 2.0 ⭐⭐⭐⭐⭐ 最佳 主力使用,其 Character Drift 已被解决
Nano Banana 2 ⭐⭐⭐⭐ 强 支持 5 角色/14 物体,适合群戏
LibLib/即梦 ⭐⭐⭐ 中等 需要更严格的 Token + Reference
其他模型 ⭐⭐ 不稳定 建议用 Seedance 处理关键角色镜头

🛠️ 工具与集成

Reference Image 使用方式

在 prompt 中引用本地文件路径:

# 使用本地路径
Reference: ./characters/protagonist/ref_front.jpg
Style Reference: ./characters/protagonist/ref_front.jpg

# 多个参考角度
Primary Ref: ./characters/protagonist/ref_front.jpg
Secondary Ref: ./characters/protagonist/ref_3quarter.jpg

推荐的 AI 工具

  • Seedance 2.0 — 目前 character consistency 最佳解决方案
  • Nano Banana 2 — 多角色/物体一致性
  • LibLib AI — 参考图 + Token 组合使用
  • 即梦 — 参考图锁定 + 负向描述

📁 相关文件

  • references/asset-preparation.md — 定妆照拍摄规范与资产准备详解
  • references/consistency-tokens.md — Consistency Tokens 完整模板与格式说明
安全使用建议
This skill is an instruction-only workflow for prompt engineering and asset organization; it does not install code or ask for credentials. Before using it, consider: (1) the workflow assumes you will reference local reference images — ensure those files are ones you own or have permission to use and do not contain sensitive personal data you don't want processed; (2) using models (Seedance, Nano Banana, etc.) may require separate API keys or services you must configure yourself — never paste secrets into prompts; (3) if you run the agent with filesystem access, be mindful that prompts referencing local paths imply the agent or the model client can read those assets; restrict agent permissions if you do not want it to access unrelated files; (4) the guidance is focused on prompting and asset management, not on legal/ethical rules for using likenesses — confirm consent and licensing for any real-person imagery.
功能分析
Type: OpenClaw Skill Name: character-consistency-workflow Version: 1.0.0 The skill bundle provides a comprehensive instructional workflow for maintaining character consistency in AI-generated media. It includes detailed guides on asset preparation (asset-preparation.md), folder organization, and the creation of descriptive 'Consistency Tokens' (consistency-tokens.md) to prevent character drift. The content is purely educational and lacks any indicators of malicious intent, data exfiltration, or unauthorized command execution.
能力标签
crypto
能力评估
Purpose & Capability
The name/description (character consistency for multi-shot AI generation) matches the content: prompt templates, token system, asset preparation and folder layout. There are no unrelated env vars, binaries, or installs requested.
Instruction Scope
The SKILL.md instructs the agent to embed local reference image paths (e.g., ./characters/...) and to reuse token blocks in prompts — this is appropriate for the stated goal. Note: it implies use of local project assets; the skill does not declare system config paths or credentials, nor does it include commands that read arbitrary system files, but an agent executing these prompts would need access to the referenced project files.
Install Mechanism
No install spec and no code files are present; the skill is instruction-only so nothing will be written to disk or downloaded by the skill itself.
Credentials
No environment variables, credentials, or external tokens are required by the skill. Model/service names are referenced (Seedance, Nano Banana, LibLib, 即梦) but providing API keys or platform credentials would be a separate, user-managed step — nothing disproportionate is requested here.
Persistence & Privilege
always is false and the skill is user-invocable; autonomous invocation is allowed by platform default but the skill does not request elevated or persistent privileges nor does it modify other skills or system-wide settings.
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install character-consistency-workflow
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /character-consistency-workflow 触发
  4. 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.0
解决人物漂移问题,三级一致性,定妆照锚定,跨镜头Token系统
元数据
Slug character-consistency-workflow
版本 1.0.0
许可证 MIT-0
累计安装 0
当前安装数 0
历史版本数 1
常见问题

Character Consistency Workflow 是什么?

AI 短剧/短片创作中的角色一致性工作流。解决"人物漂移"(character drift)问题——多镜头生成时同一角色外观漂移、换脸、换装。触发关键词:角色一致性、人物漂移、换脸、定妆照、AI视频、多镜头、一致性Tokens、Seedance、Nano Banana。当需要保持角色跨镜头外观统一、防止AI生成时... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 40 次。

如何安装 Character Consistency Workflow?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install character-consistency-workflow」即可一键安装,无需额外配置。

Character Consistency Workflow 是免费的吗?

是的,Character Consistency Workflow 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。

Character Consistency Workflow 支持哪些平台?

Character Consistency Workflow 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。

谁开发了 Character Consistency Workflow?

由 beermanzz(@beermanzz)开发并维护,当前版本 v1.0.0。

💬 留言讨论