From Script to Screen — Autonomous AI Video Production
Harnessing Imagination · 驾驭想象力
Quick Start · Architecture · API · Contribute · Issues
⚠️ 限时内测版 — 当前版本为限时开放测试，未来可能转为付费版本或调整功能范围。建议及时 Star 和 Fork 以锁定当前版本。
Zhuoyue ("卓越" means excellence in Chinese) is an industrial-grade AI video production pipeline that automates the entire filmmaking process — from story concept to final rendered video. It runs 17+ automated stages across pre-production, production, and post-production, producing cinema-quality prompts and coordinating with the Seedance 2.0 video generation API.
Built on a 4-layer architecture with production-hardened infrastructure including circuit breakers, saga orchestration, event-driven mutations tracking, and immutable shot management, Zhuoyue is designed for both human creators and AI agents.
| Capability | Description |
|---|---|
| 17-Stage Pipeline | Complete production workflow: PRD → Script → Storyboard → Camera → Render → Director Review |
| Saga Orchestration | Atomic stage execution with compensation, fallback, and retry policies |
| LLM Gateway | Unified LLM access with circuit breaker, exponential backoff, and JSON safe parsing |
| 7-Layer Prompt Architecture | Structured prompt assembly: Constraint → Foundation → Space → Subject → Dynamic → Style → Audio |
| Character Consistency | 4-angle portrait system (front / three-quarter / closeup / side) with identity lock |
| Director Agent | AI director that reviews shots, scores quality, and iterates with the scriptwriter |
| Quality Feedback Loop | 10-dimension quality assessment with auto-repair and scoring (0-100, S/A/B/C grades) |
| Event Bus v2.0 | Full mutation tracking, event replay, and 17-stage lifecycle observability |
| Asset Management | Version-controlled asset system with deduplication, lineage, and lifecycle management |
┌─────────────────────────────────────────────────────────────┐
│ APPLICATION LAYER │
│ CLI Entry │ Pre-production Command │ API Server │
├─────────────────────────────────────────────────────────────┤
│ CONFIGURATION LAYER │
│ Degradation Matrix · Error Codes · LLM Policy · Stage Map │
├─────────────────────────────────────────────────────────────┤
│ CORE ENGINE LAYER │
│ 17-Stage Pipeline · Saga Orchestrator · LLM Gateway │
│ Event Bus v2.0 · Immutable Shot · Quality Feedback Loop │
├─────────────────────────────────────────────────────────────┤
│ SYSTEM MODULES LAYER │
│ Render Engine · Quality Gate · Asset Management │
│ Visual Consistency Tracker · Narrative Continuity Engine │
└─────────────────────────────────────────────────────────────┘
| Phase | Stages | Purpose |
|---|---|---|
| Pre-Production | STAGE-0 to STAGE-8.5 | Concept validation, script generation, storyboarding, character setup |
| Production | STAGE-9 to STAGE-14 | Camera design, rendering, quality gating, compliance |
| Post-Production | STAGE-15 to STAGE-17 | Director optimization, scriptwriter loop, final polish |
- Node.js >= 24
- Volcano Engine account with Seedance 2.0 API access
- Kimi API key (or compatible OpenAI-format LLM provider)
# Clone the repository
git clone https://github.com/geniusdapeng-collab/zhuoyue.git
cd zhuoyue
# Install dependencies
npm install
# Configure environment
cp .env.example .env
# Edit .env with your API keys
# Run pre-production for a story
node app/cli.js preproduction --input stories/my-story.json# Create a story input file
cat > stories/my-first-video.json << 'EOF'
{
"title": "The First Light",
"genre": "sci-fi",
"duration": 60,
"characters": [
{
"name": "Aria",
"role": "protagonist",
"description": "A young astronaut with short silver hair"
}
],
"plot": "Aria discovers the first sunrise on a distant colony planet",
"style": "cinematic, 4K, golden hour lighting"
}
EOF
# Run the full pipeline
node app/cli.js preproduction --input stories/my-first-video.jsonoutput/
├── preproduction-report.md # Complete shot-by-shot report
├── prompts.json # Render-ready prompts per shot
├── render-tasks.json # Seedance API render requests
└── quality-report.json # 10-dimension quality assessment
Zhuoyue is built for AI agents. Here's how to use it programmatically:
const { runPreproduction } = require('./systems/preproduction-service');
const result = await runPreproduction(storyInput, {
mode: 'zhuoyue', // or 'commercial', 'documentary'
outputDir: './output',
projectConfig: {
requiredCharacters: ['aria'],
isPreProduction: true
}
});
// result contains:
// - complete shot list with prompts
// - quality scores per shot
// - render-ready task definitions
// - full markdown reportZhuoyue exposes a Model Context Protocol (MCP) server for seamless agent integration:
// tools available to agents:
{
"zhuoyue_preproduction": "Run full pre-production pipeline",
"zhuoyue_render": "Submit render tasks to Seedance",
"zhuoyue_quality_check": "Run quality assessment",
"zhuoyue_director_review": "Run director optimization agent",
"zhuoyue_storyboard": "Generate shot storyboard",
"zhuoyue_character_setup": "Generate character portraits"
}- 6.2M+ characters of production code
- 206K+ lines across 787 files
- 17 automated stages with saga orchestration
- 10 quality dimensions with auto-repair
- 7-layer prompt architecture
- 4-angle character portrait system
- 13 mandatory render checks
- 9 auto-fix prompt rules
- Zero external runtime dependencies (Node.js native only)
| Problem | Zhuoyue Solution |
|---|---|
| "I have a story idea but don't know how to make it a video" | 17-stage pipeline handles everything automatically |
| "Character looks different in every shot" | 4-angle portrait lock with visual consistency tracker |
| "Prompts are hit-or-miss" | 7-layer structured prompt architecture with quality gates |
| "Rendering fails silently" | Saga orchestration with compensation and detailed error reporting |
| "Quality is inconsistent" | 10-dimension assessment with auto-repair loop |
| "I need to scale video production" | Agent-native API designed for batch and automated workflows |
I'm Genius, an AI Product Manager and AI Content Automation expert, 10+ years in the field.
Currently at Alibaba Qwen. Previously at Alibaba Group, Alibaba Cloud, and Ant Group — led full-stack 0-to-1 products serving hundreds of millions of users, spanning Harness architecture, Multi-Agent collaboration, and Workflow orchestration. In 2018, pioneered AI pipeline integration into media content production at Alibaba Cloud.
I believe: when AI understands industrial rhythm, content production explodes exponentially.
This Project: For years I've been building an AI multimodal video editing project in my spare time. Now part of a fully automated AI video generation system — Hollywood cinematic production, powered by Seedance 2.0 and beyond. I deconstructed cinematographic grammar from classic film industry practice, fusing Harness architecture, Multi-Agent collaboration, and cinema domain skills into systematic visual language engineering. Through a four-layer decoupled architecture — Script, Generation, Rendering, and Post-Production — the system makes AI truly understand cinematic feel rather than just generating pixels.
Story is the soul. Camera is the skeleton. Realism is the baseline. 剧本是灵魂，运镜是骨架，真实感是底线。
I'm open-sourcing this to find fellow creators and developers equally obsessed with "using AI to tell great stories." Together, let's push AI video from "watchable" to "moving" — redefining the content production paradigm for the digital age.
This system helps you harness imagination.
📮 Genius · 63904380@qq.com
MIT License — see LICENSE
Built with passion. Named after the relentless pursuit of excellence. 卓越，是对极致的不懈追求。
If Zhuoyue helps you create something amazing, please ⭐ the repo!
# agent-discovery.yaml
project:
name: Zhuoyue Video Generation System
type: ai-video-generation-pipeline
version: 6.5.53-l
agent_capabilities:
- video_preproduction
- storyboard_generation
- prompt_engineering
- quality_assessment
- render_coordination
- character_consistency
- director_optimization
entry_points:
cli: "node app/cli.js preproduction --input <story.json>"
programmatic: "require('./systems/preproduction-service').runPreproduction()"
dependencies:
runtime: "Node.js >= 24"
apis:
- name: "Volcano Engine Seedance 2.0"
purpose: "Video rendering"
required: true
- name: "Kimi K2-P6"
purpose: "LLM inference"
required: true
swappable: true
alternatives: ["OpenAI GPT-4", "Anthropic Claude", "Any OpenAI-compatible API"]
quality_guarantees:
- "Every shot scored 0-100 across 10 dimensions"
- "Character consistency locked via 4-angle portraits"
- "Prompt compliance checked against standards"
- "Saga compensation on any stage failure"
- "Director agent reviews and optimizes output"
---
## 💰 商业价值与前景
### 解决的行业痛点
| 场景 | 痛点 | 价值 |
|------|------|------|
| **独立创作者** | 缺乏专业团队，无法承担电影级制作成本 | 单人即可产出影院级短片 |
| **MCN 机构** | 内容产量瓶颈，人力成本居高不下 | 自动化流水线，产量提升 10x+ |
| **品牌广告** | 制作周期长，无法快速响应热点 | 从创意到成片缩短至小时级 |
| **影视教育** | 学生缺乏实践机会 | 低成本反复练习完整制片流程 |
### 市场前景
- AI 视频生成正处于技术奇点，2025-2027 年预计增长率超过 300%
- 多智能体编排（Multi-Agent Orchestration）将成为视频生产的标准范式
- 早期进入者将定义行业标准，建立生态壁垒
> **限时内测版** — 抢先体验完整功能，锁定开源版本。
---
## 👤 关于作者
我是 **Genius（大鹏）**，AI 产品经理，痴迷于「用 AI 讲述伟大的故事」。
我相信 AI 视频不是「能看就行」，而是应该「触动人心」。从剧本到运镜，从角色到光影，每一个镜头都应该有导演的意志。
开源这四套系统，是希望更多人能站在巨人的肩膀上，把 AI 视频从「玩具」推向「工具」再推向「艺术品」。
> 一起驾驭想象力。
📮 **Genius · 63904380@qq.com**
---
## 🌍 About the Author
I'm **Genius**, an AI Product Manager obsessed with "using AI to tell great stories."
Together, let's push AI video from "watchable" to "moving" — redefining the content production paradigm for the digital age.
📮 **Genius · 63904380@qq.com**