AI视频生成

Setup

On first use, read setup.md.

When to Use

User needs to generate, edit, or scale AI videos with current models and APIs.
Use this skill to choose the right current model stack, write stronger motion prompts, and run reliable async video pipelines.

Architecture

User preferences persist in ~/video-generation/. See memory-template.md for setup.

CODEBLOCK0

Quick Reference

Topic	File
Initial setup	INLINECODE3
Memory template

memory-template.md | | Migration guide | migration.md | | Model snapshot | benchmarks.md | | Async API patterns | api-patterns.md | | OpenAI Sora 2 | openai-sora.md | | Google Veo 3.x | google-veo.md | | Runway Gen-4 | runway.md | | Luma Ray | luma.md | | ByteDance Seedance | seedance.md | | Kling | kling.md | | Vidu | vidu.md | | Pika via Fal | pika.md | | MiniMax Hailuo | minimax-hailuo.md | | Replicate routing | replicate.md | | Open-source local models | open-source-video.md | | Distribution playbook | promotion.md |

Core Rules

1. Resolve model aliases before API calls

Map community names to real API model IDs first.
Examples: sora-2, sora-2-pro, veo-3.0-generate-001, gen4_turbo, gen4_aleph.

2. Route by task, not brand preference

Task	First choice	Backup
Premium prompt-only generation	INLINECODE25	INLINECODE26
Fast drafts at lower cost

veo-3.1-fast-generate-001 | gen4_turbo | | Long-form cinematic shots | gen4_aleph | ray-2 | | Strong image-to-video control | veo-3.0-generate-001 | gen4_turbo | | Multi-shot narrative consistency | Seedance family | hailuo-2.3 | | Local privacy-first workflows | Wan2.2 / HunyuanVideo | CogVideoX |

3. Draft cheap, finish expensive

Start with low duration and lower tier, validate motion and composition, then rerender winners with premium models or longer durations.

4. Design prompts as shot instructions

Always include subject, action, camera motion, lens style, lighting, and scene timing.
For references and start/end frames, keep continuity constraints explicit.

5. Assume async and failure by default

Every provider pipeline must support queued jobs, polling/backoff, retries, cancellation, and signed-URL download before expiry.

6. Keep a fallback chain

If the preferred model is blocked or overloaded:
1) same provider lower tier, 2) equivalent cross-provider model, 3) open model/local run.

Common Traps

• - Using nickname-only model labels in code -> avoidable API failures
• Pushing 8-10 second generations before validating a 3-5 second draft -> wasted credits
• Cropping after generation instead of generating native ratio -> lower composition quality
• Ignoring prompt enhancement toggles -> tone drift across providers
• Reusing expired output URLs -> broken export workflows
• Treating all providers as synchronous -> stalled jobs and bad timeout handling

External Endpoints

Provider	Endpoint	Data Sent	Purpose
OpenAI	INLINECODE34	Prompt text, optional input images/video refs	Sora 2 video generation
Google Vertex AI

aiplatform.googleapis.com | Prompt text, optional image input, generation params | Veo 3.x generation | | Runway | api.dev.runwayml.com | Prompt text, optional input media | Gen-4 generation and image-to-video | | Luma | api.lumalabs.ai | Prompt text, optional keyframes/start-end images | Ray generation | | Fal | queue.fal.run | Prompt text, optional input media | Pika and Hailuo hosted APIs | | Replicate | api.replicate.com | Prompt text, optional input media | Multi-model routing and experimentation | | Vidu | api.vidu.com | Prompt text, optional start/end/reference images | Vidu text/image/reference video APIs | | Tencent MPS | mps.tencentcloudapi.com | Prompt text and generation parameters | Unified AIGC video task APIs |

No other data is sent externally.

Security & Privacy

Data that leaves your machine:

• - Prompt text
• Optional reference images or clips
• Requested rendering parameters (duration, resolution, aspect ratio)

Data that stays local:

• - Provider preferences in INLINECODE42
• Optional local job history in INLINECODE43

This skill does NOT:

• - Store API keys in project files
• Upload media outside requested provider calls
• Delete local assets unless the user asks

Trust

This skill can send prompts and media references to third-party AI providers.
Only install if you trust those providers with your content.

Related Skills

Install with clawhub install <slug> if user confirms:

• - image-generation - Build still concepts and keyframes before video generation
• INLINECODE46 - Prepare clean references, masks, and style frames
• INLINECODE47 - Post-process generated clips and final exports
• INLINECODE48 - Add subtitle and text overlay workflows
• INLINECODE49 - Compose, transcode, and package production outputs

Feedback

• - If useful: INLINECODE50
• Stay updated: INLINECODE51