← 返回 Skills 市场
aaalenwow

Ai Video Pro

作者 AAAlenwow · GitHub ↗ · v0.2.0 · MIT-0
cross-platform ⚠ suspicious
352
总下载
0
收藏
2
当前安装
2
版本数
在 OpenClaw 中安装
/install ai-video-pro
功能描述
电影级镜头语言解码器。零依赖可用:将创意描述解码为影视级提示词,系统化建模打击感/面部表情渐变/机甲运动五维镜头语言。可选配 API Key 直接生成视频(LumaAI/Runway/Replicate/ComfyUI)。提示词优化本身无需任何外部服务。
使用说明 (SKILL.md)

用户提供创意描述,输出影视级提示词。可选连接生成后端直接产出视频。

⚠️ BETA — 勿用于生产环境。


运行模式

模式 依赖 输出
提示词模式(默认) 影视级优化 prompt,可复制到任意生成工具
生成模式 API Key 或本地 GPU 直接调用后端生成视频文件

提示词模式是核心价值所在:无需任何配置,立即可用,输出的 prompt 可直接用于 LumaAI / Runway / ComfyUI 等任意工具。


调用示例

# 基础描述 → 影视级 prompt
帮我把这个描述转成视频 prompt:一个武士在雨中拔刀

# 指定风格和平台
生成一个抖音竖屏的动作场景 prompt,风格写实,要有打击感

# 迭代优化
上一个 prompt 的打击感不够,帮我加强一下

# 生成模式(需配置 API Key)
用 LumaAI 生成这个场景:[场景描述]

核心方法论:五维镜头语言解码

用户描述 ──解码──→ [镜头类型 × 运镜 × 灯光 × 色彩 × 时间] ──适配──→ Provider Prompt

普通描述和电影级描述的差异在于隐性专业知识的显性化

维度 解码内容 关键判断点
镜头类型 ECU/CU/MS/FS/WS/鸟瞰/仰拍/荷兰角 情感距离与叙事权力
运镜方式 固定/Pan/Dolly/Tracking/手持/斯坦尼康 观众与场景的关系感
灯光设计 轮廓光/伦勃朗/体积光/黄金时刻/霓虹 情绪基调与质感
色彩分级 青橙对比/去饱和/胶片模拟/冷暖对比 风格化程度
时间控制 慢动作/延时/变速/定格 叙事节奏

Phase 1: 镜头语言解码(提示词模式,零依赖)

1.1 动作/冲击场景 — 打击感五要素

国内短视频最核心的差异化维度,必须精确建模:

① 重量感    — 攻击者的预备动作幅度、身体质量感
② 力度反馈  — 被击中者的形变程度(衣物/头发/身体弯曲)
③ 冲击特效  — 火花/碎片/冲击波的密度和扩散方向
④ 时间节奏  — 预备→接触→反应的帧率变化(慢进快出 / 快进慢出)
⑤ 画面反应  — 镜头抖动幅度、运动模糊方向、色差效果

未指定时询问:打击力度(轻触/重击/影视夸张)、特效风格(写实/漫画/机甲)。

1.2 角色场景 — 面部表情渐变序列

建模表情的时间序列变化,而非静态描述:

示例:[0s] 惊讶(眉上扬、瞳孔扩张)→ [2s] 坚定(下颌收紧、目光聚焦)→ [5s] 胜利(嘴角微扬)

未指定时询问:镜头起止的表情分别是什么、是否有眼神交流。

1.3 机甲/硬核运动

额外建模:关节铰接方向、液压延迟感、质量惯性(重型机甲响应慢)、变形序列节点。

1.4 缺失维度补全

解码时主动检查必要参数,如未指定则询问:

  • 画面比例(16:9 横屏 / 9:16 抖音竖屏 / 1:1 小红书)
  • 目标时长(3s/5s/10s)
  • 视觉风格(写实/动漫/3D/水彩)

1.5 Provider 适配输出

根据目标平台重新编码 prompt 风格(或同时输出多平台版本):

Provider Prompt 偏好
LumaAI Dream Machine 自然语言嵌入镜头指令,"camera slowly pans…"
Runway Gen-3/4 结构化,镜头/主体/风格分段
Replicate/SVD 简洁,强调主体动态
ComfyUI 本地 checkpoint + scheduler + LoRA 标签

同时展示原始描述和优化后 prompt,用户确认后再进行下一步。


Phase 2: 视频生成(生成模式,需 API Key 或 GPU)

后端选择(最小代价优先):

优先级 后端 条件 成本
1 ComfyUI 本地 NVIDIA 8GB+ VRAM 免费
2 Replicate 免费层 API Token 免费(有限)
3 LumaAI 免费层 API Key 免费(有限)
4 LumaAI 付费 API Key ~¥3.5/视频
5 Runway 付费 API Key ~¥7/视频

推荐方案和预估成本先展示给用户确认,再调用 API。

python3 scripts/env_detect.py          # 检测可用后端
python3 scripts/provider_manager.py --backend \x3Cbackend> --prompt \x3Cprompt_file>
python3 scripts/preview_server.py --file \x3Cvideo_path> --port 8765

数据流向:优化后的 prompt 发送至用户选择的生成后端(LumaAI/Runway/Replicate)。原始描述仅在本次会话中处理,不向其他第三方传输。


Phase 3: 迭代优化

保留解码结构,仅调整目标维度:

反馈 调整方向
打击感不足 增加特效密度、加强时间节奏对比
镜头感弱 升级运镜(手持 → 斯坦尼康 / 固定 → Dolly)
情绪不对 调整色彩分级基调 + 表情序列终点
风格漂移 固定 LoRA/风格标签,切换 Provider

支持图生视频:上传关键帧作为风格锚点,保持角色一致性。


输出格式(提示词模式)

{
  "original_description": "用户原始描述",
  "decoded_dimensions": {
    "shot_type": "CU 近景",
    "camera_movement": "手持跟拍",
    "lighting": "伦勃朗光 + 轮廓光",
    "color_grading": "青橙对比,高饱和",
    "temporal": "接触瞬间慢动作 0.3x"
  },
  "prompts": {
    "lumai": "A close-up shot, handheld camera following...",
    "runway": "Shot type: CU. Subject: ...",
    "universal": "通用版本(可用于任意平台)"
  },
  "missing_params": ["画面比例未指定,建议确认"]
}

边界与合规

本技能做什么: 提示词工程优化、镜头语言建模、生成后端调度。

本技能不做什么:

  • 提示词模式不需要、不访问任何外部服务
  • 不存储用户提交的创意描述
  • 不自动发布到任何平台(发布需要用户显式授权和凭证配置)

凭证(均为可选,仅生成模式需要):

  • LUMAAI_API_KEY / RUNWAY_API_KEY / REPLICATE_API_TOKEN / OPENAI_API_KEY
  • 所有凭证仅通过环境变量读取,不持久化、不记录
安全使用建议
What to consider before installing/using this skill: - Functionality & scope: The prompt-optimization feature is local and usable without credentials; generation/publishing features require optional API keys and cloud/platform credentials. Only provide keys you actually need (e.g., LUMAAI_API_KEY only if you plan to call LumaAI). - Metadata mismatch: The skill's registry metadata lists no required env vars, but the code references many optional credentials. Treat those optional creds as sensitive — don't set them globally unless you intend to use generation/publishing. - Inspect before running: Review provider_manager.py and publish.py (not fully shown in the manifest) to confirm where and how credentials are used and whether any unexpected endpoints appear. - Be cautious with install_deps.py: it runs package managers, pip installs and can clone ComfyUI from GitHub. Only run it in an environment you control (VM, container, or isolated machine). Prefer manual dependency installation so you understand what changes are made. - Publishing risks: Xiaohongshu support warns that cookie-based automation may violate TOS. Avoid supplying XHS_COOKIE unless you understand and accept that risk. - Minimize credential exposure: Use per-skill, least-privilege API keys if possible. Do not paste long-lived cloud credentials in environments used for unrelated tasks. - Run harmless checks first: Use prompt-only flows and the provided env_detect.py (it only detects presence/absence and probes hardware/network) to understand behavior. Do not run generation/publish steps until you’re comfortable. - If unsure: run the skill in an isolated VM/container, or ask someone with system-admin experience to review the code paths you plan to use. If you want, I can: (a) list the exact env vars the code reads, (b) scan provider_manager.py and publish.py for network endpoints and calls, or (c) produce a short checklist of commands to safely try prompt-only mode without installing dependencies.
功能分析
Type: OpenClaw Skill Name: ai-video-pro Version: 0.2.0 The ai-video-pro skill bundle is a comprehensive tool for cinematic video prompt engineering and multi-backend video generation. While it possesses high-risk capabilities—including managing numerous sensitive API keys (LumaAI, Runway, Weibo, AWS, etc.), executing shell commands for video processing (ffmpeg) and dependency installation, and running a local HTTP server for previews—these actions are well-documented and strictly aligned with the stated functionality. The code demonstrates good security practices, such as a credential manager that explicitly avoids logging or persisting secrets (credential_manager.py) and a transparent environment detection script (env_detect.py). No evidence of data exfiltration, unauthorized remote control, or malicious prompt injection was found.
能力评估
Purpose & Capability
The skill claims to be primarily a 'prompt optimization' tool with optional generation. That matches included files (prompt_engine.py, prompt templates). However the registry metadata (requires.env: [] / primaryEnv: LUMAAI_API_KEY) understates the number of optional credentials and integrations present in code: credential_manager.py, env_detect.py, provider_manager.py and platform configs reference many other env vars (RUNWAY_API_KEY, REPLICATE_API_TOKEN, OPENAI_API_KEY, DOUYIN_ACCESS_TOKEN, XHS_COOKIE, cloud provider keys, etc.). The presence of publish.py and platform config files shows full publish/upload capability (including cookie-based publishing for XHS), which is coherent with README but broadens the skill's scope beyond pure local prompt optimization.
Instruction Scope
SKILL.md describes a safe 'prompt-only' default mode and an optional 'generation' mode requiring API keys or GPU. The runtime instructions call scripts (env_detect.py, provider_manager.py, preview_server.py). env_detect probes local hardware, installed tools, disk and network (it performs network checks). credential_manager reads environment variables for many providers. The SKILL.md asserts '提示词模式不访问任何外部服务' — that appears true if you only use prompt-mode, but the shipped scripts will access external services if you run generation/publish flows. The agent instructions and code give the agent the capability to (with explicit user action) run commands that access network and system state.
Install Mechanism
There is no automatic install spec in registry (instruction-only), but the repository contains install_deps.py which, when executed, runs system package managers (winget/brew/apt), pip installs, and can clone ComfyUI from GitHub. These operations are standard for enabling local generation but are system-level and will write to disk and install third-party code. The ComfyUI clone source is a public GitHub repository (comfyanonymous/ComfyUI), not an arbitrary URL shortener or IP, which reduces risk but still requires user review and consent before running.
Credentials
The skill declares LUMAAI_API_KEY as primaryEnv but code expects and documents many optional credentials (LUMAAI_API_KEY, RUNWAY_API_KEY, REPLICATE_API_TOKEN, OPENAI_API_KEY, KLING_API_KEY, WEIBO_ACCESS_TOKEN, XHS_COOKIE, DOUYIN_ACCESS_TOKEN, plus cloud keys for Aliyun/Tencent/AWS). None are required for prompt-mode, but the presence of cloud storage and platform upload keys (AWS/ALIYUN/TENCENT) is broad relative to the core purpose (prompt optimization). The metadata's empty requires.env is therefore incomplete. The code claims not to persist or print secrets (credential_manager enforces in-code guidance), but giving cloud/platform credentials enables upload/publish features and increases blast radius if misused.
Persistence & Privilege
The skill is not always-enabled, does not request permanent platform-wide privileges, and the code states it will not persist credentials to disk. The preview server binds only to localhost. There is no manifest 'always: true' and no evidence the skill autonomously modifies other skills. However some scripts (install_deps.py, install_comfyui) create files on disk and clone projects if you run them — this is expected install behavior but should be consciously executed by the user.
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 ai-video-pro
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /ai-video-pro 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
v0.2.0
优化:AI推理优先架构(零依赖可用);添加运行模式/调用示例/数据流向/边界声明;py脚本增强:否定语境检测、滑动窗口检测、镜头元素提取修复、复合故障矩阵、角色弧线追踪
v0.1.0
Initial beta release - cinematic prompt optimization, auto-detection of video generation backends, multi-platform publishing. NOT for production use.
元数据
Slug ai-video-pro
版本 0.2.0
许可证 MIT-0
累计安装 2
当前安装数 2
历史版本数 2
常见问题

Ai Video Pro 是什么?

电影级镜头语言解码器。零依赖可用:将创意描述解码为影视级提示词,系统化建模打击感/面部表情渐变/机甲运动五维镜头语言。可选配 API Key 直接生成视频(LumaAI/Runway/Replicate/ComfyUI)。提示词优化本身无需任何外部服务。 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 352 次。

如何安装 Ai Video Pro?

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

Ai Video Pro 是免费的吗?

是的,Ai Video Pro 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。

Ai Video Pro 支持哪些平台?

Ai Video Pro 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 Ai Video Pro?

由 AAAlenwow(@aaalenwow)开发并维护,当前版本 v0.2.0。

💬 留言讨论