一个兼容 Codex 与 Claude Code 的专业图生视频运镜 Skill。它会先判断静态图片能够安全支持多大幅度的镜头运动，再根据画面主体、商业目标和模型能力，设计推拉摇移、升降、环绕、跟拍、微距滑动等镜头，并输出可直接交给图生视频模型使用的提示词与结构化运动方案。
核心能力采用通用的 SKILL.md + references + scripts 目录结构，Codex 和 Claude Code 均可读取同一套技能文件，无需维护平台专用副本。
| 平台 | 安装位置 | 显式调用方式 |
|---|---|---|
| Codex | $CODEX_HOME/skills/image-to-video-director/，未设置时通常为 ~/.codex/skills/image-to-video-director/ |
$image-to-video-director |
| Claude Code（个人） | ~/.claude/skills/image-to-video-director/ |
/image-to-video-director |
| Claude Code（项目） | <项目目录>/.claude/skills/image-to-video-director/ |
/image-to-video-director |
Claude Code 会使用 SKILL.md、references/ 和 scripts/。agents/openai.yaml 仅用于 Codex 的技能列表与默认提示界面，Claude Code 会忽略该文件，不影响技能执行。两个平台也都可以根据 SKILL.md 中的描述，在符合场景时自动选择该技能。
- 识别产品、人物、建筑、界面、食物、环境等主体类型，并确认所需素材与制作路径。
- 分析景深、遮挡、边缘余量、刚性结构、文字和身份一致性等风险。
- 区分平移、旋转、变焦、环绕和跟拍，控制运动幅度、速度、节奏与结束画面。
- 根据不同模型的实际能力适配提示词、负面提示词、时长、画幅和相机控制参数。
- 对生成结果进行身份、几何、运动连续性、焦点和收尾质量检查，并给出修正方案。
- 通过确定性的 Python 校验器检查运动计划结构和常见高风险组合。
- 电商产品展示和品牌广告镜头
- 人像、角色或时尚素材的轻量动态化
- 建筑、室内与风景图片的空间展示
- 食品、美妆、珠宝等细节特写
- 网站和软件界面的确定性平移、缩放或合成方案
- 已生成图生视频结果的失真诊断与提示词修复
Skill 默认遵循以下流程：
- 识别主体类型、证据、所需素材和建议制作路径，并等待用户确认。
- 检查源图能否支撑目标运镜，确定最大安全运动幅度。
- 明确镜头目的，选择一个主导运动并设计起止画面。
- 输出中文导演说明、英文模型提示词和结构化运动计划。
- 在存在已授权生成工具时适配模型、生成视频并进行视觉质检。
对于单张图片无法可靠重建的内容，Skill 会主动降低运动幅度。例如，将 360 度环绕改为小角度弧线移动，或建议补充多视角素材，而不是用提示词强行生成不可见表面。
将整个仓库克隆或复制到对应平台的技能目录，并确保目录名为 image-to-video-director。安装后重新启动会话，使平台重新发现技能。
Codex 个人安装示例：
git clone https://github.com/HiAPIAI/hiapi-image-to-video-camera-motion-skill.git ~/.codex/skills/image-to-video-directorClaude Code 个人安装示例：
git clone https://github.com/HiAPIAI/hiapi-image-to-video-camera-motion-skill.git ~/.claude/skills/image-to-video-directorClaude Code 项目级安装时，将目标目录改为当前项目下的 .claude/skills/image-to-video-director。项目级技能可随项目提交并由团队共享。
运行 scripts/validate_motion_plan.py 需要 Python 3.9 或更高版本。
在 Codex 中可通过 $image-to-video-director 显式调用：
使用 $image-to-video-director 分析这张产品图，设计一个 5 秒的缓慢推近镜头，保持包装文字和 Logo 不变。
使用 $image-to-video-director 检查这段图生视频为什么人物面部和背景在环绕时变形，并给出更稳妥的重做方案。
在 Claude Code 中使用 /image-to-video-director：
/image-to-video-director 分析这张建筑图片，设计一个安全的小幅向右弧线运镜，并输出模型可用的英文提示词。
.
├── SKILL.md # Skill 入口、工作流与输出契约
├── agents/openai.yaml # Codex 界面元数据，Claude Code 可忽略
├── references/ # 运镜、源图分析、模型路由与质检资料
└── scripts/validate_motion_plan.py # 结构化运动计划校验器
将结构化计划保存为 JSON 后运行：
python scripts/validate_motion_plan.py plan.json严格模式会把警告也视为失败：
python scripts/validate_motion_plan.py plan.json --strict校验器负责检查字段完整性和高风险运镜组合，不能替代模型参数校验或对实际视频的视觉审查。
- 一条短视频只保留一个主导镜头运动。
- 优先保护主体身份、产品结构、文字、Logo、光线和空间关系。
- 把大幅平移与环绕视为重建任务，而不是普通风格词。
- 优先生成保守基线，每次只改变一个变量。
- 没有生成或检查工具时，不声称已经生成或通过视觉验收。
A professional image-to-video camera direction skill compatible with Codex and Claude Code. It first determines how much camera movement a still image can safely support, then designs push-ins, pull-outs, pans, tilts, pedestal and crane moves, arcs, orbits, tracking shots, and macro glides based on the subject, production goal, and actual model capabilities. The skill produces model-ready prompts and structured motion plans.
The core capability uses a portable SKILL.md + references + scripts structure. Codex and Claude Code can load the same skill files without separate platform-specific implementations.
| Platform | Installation location | Explicit invocation |
|---|---|---|
| Codex | $CODEX_HOME/skills/image-to-video-director/, usually ~/.codex/skills/image-to-video-director/ when unset |
$image-to-video-director |
| Claude Code (personal) | ~/.claude/skills/image-to-video-director/ |
/image-to-video-director |
| Claude Code (project) | <project>/.claude/skills/image-to-video-director/ |
/image-to-video-director |
Claude Code uses SKILL.md, references/, and scripts/. The agents/openai.yaml file only provides Codex skill-list and default-prompt metadata; Claude Code ignores it without affecting skill behavior. Both platforms may also select the skill automatically when a request matches the description in SKILL.md.
- Classify products, people, architecture, interfaces, food, environments, and other subjects, then confirm the required assets and production path.
- Analyze depth, occlusion, edge reserve, rigid geometry, typography, and identity-preservation risks.
- Distinguish camera translation, rotation, zoom, orbit, and tracking, with bounded travel, speed, timing, and end-frame intent.
- Adapt prompts, negative prompts, duration, aspect ratio, and camera controls to the capabilities actually exposed by a selected model.
- Review generated clips for identity, geometry, motion continuity, focus, and end-frame quality, then propose targeted corrections.
- Validate structured motion plans and common high-risk combinations with a deterministic Python script.
- E-commerce product showcases and brand advertising shots
- Subtle animation for portraits, characters, and fashion imagery
- Spatial presentation of architecture, interiors, and landscapes
- Detail shots for food, beauty products, jewelry, and similar subjects
- Deterministic pan, zoom, scroll, or compositing plans for websites and software interfaces
- Distortion diagnosis and prompt repair for existing image-to-video results
The skill follows this default process:
- Classify the subject, supporting evidence, required assets, and proposed production path, then wait for user confirmation.
- Inspect whether the source image can support the requested move and establish the maximum safe movement.
- Define the shot purpose, select one dominant camera behavior, and design the start and end frames.
- Produce a Chinese director note, an English model-ready prompt, and a structured motion plan.
- When an authorized generation tool is available, adapt the plan to the model, generate the clip, and perform visual quality control.
When a single image cannot reliably support the requested reconstruction, the skill reduces the movement to a viable alternative. For example, it may replace a 360-degree orbit with a small inspection arc or request additional views instead of forcing a model to invent unseen surfaces.
Clone or copy the entire repository into the appropriate skill directory and name the destination folder image-to-video-director. Start a new session after installation so the platform can discover the skill.
Codex personal installation:
git clone https://github.com/HiAPIAI/hiapi-image-to-video-camera-motion-skill.git ~/.codex/skills/image-to-video-directorClaude Code personal installation:
git clone https://github.com/HiAPIAI/hiapi-image-to-video-camera-motion-skill.git ~/.claude/skills/image-to-video-directorFor a project-level Claude Code installation, use .claude/skills/image-to-video-director inside the target project. Project-level skills can be committed and shared with the team.
Python 3.9 or later is required to run scripts/validate_motion_plan.py.
Invoke the skill explicitly in Codex with $image-to-video-director:
Use $image-to-video-director to analyze this product image and design a slow five-second push-in while preserving the packaging text and logo.
Use $image-to-video-director to diagnose why the face and background deform during the orbit in this generated clip, then propose a safer revision.
Invoke it in Claude Code with /image-to-video-director:
/image-to-video-director Analyze this architecture image, design a safe small arc to the right, and produce an English model-ready prompt.
.
|-- SKILL.md # Skill entry point, workflow, and output contract
|-- agents/openai.yaml # Codex UI metadata; ignored by Claude Code
|-- references/ # Motion, source analysis, model routing, and QA guidance
`-- scripts/validate_motion_plan.py # Structured motion-plan validator
Save the structured plan as JSON and run:
python scripts/validate_motion_plan.py plan.jsonStrict mode treats warnings as failures:
python scripts/validate_motion_plan.py plan.json --strictThe validator checks field completeness and high-risk motion combinations. It does not replace provider-specific schema validation or visual review of the generated video.
- Keep one dominant camera behavior in each short clip.
- Preserve subject identity, product geometry, text, logos, lighting, and spatial relationships.
- Treat large translations and orbits as reconstruction requests, not decorative style terms.
- Generate a conservative baseline first and change one variable per iteration.
- Never claim that a clip was generated or visually approved when generation or inspection tools were unavailable.