zcbacxc/movie-narrator

One Prompt → One Narrated Movie Video

Movie Narrator is an open-source toolkit that automatically generates movie recap videos with narration, subtitles, and rendered output from a simple command.

• 🎬 LLM-powered movie recap script generation
• 🔊 Text-to-Speech narration (Edge-TTS by default)
• 💬 Automatic SRT subtitle generation
• 🌐 Multi-language subtitles with LLM translation
• 🏁 Multi-candidate horse race — run N variations, auto-pick the best
• 🎯 Reference video imitation — extract style from viral narration
• 👁️ VLM scene captioning via cloud VLM API
• 🎭 Narrator perspective (omniscient / character / detective)
• 🎨 Render template system (title cards, watermarks, slogans)
• 🔍 TMDB fact verification with source attribution
• 🖥️ Web UI (separate movie-narrator-web package)
• 🎞️ Video rendering (1080p/4K output)
• 📝 Script markdown export
• 🎵 Background music integration
• 🔌 Extensible plugin architecture (custom TTS / LLM backends)
• ☁️ Async task processing (submit and go, notified on completion)
• 🌐 Remote inference via REST API
• ✅ Final-video QA — black-frame / slideshow-risk detection
• 🗣️ Optional Chinese ASR (automatic multi-backend fallback)

• Python 3.10+

pip install movie-narrator

git clone https://github.com/zcbacxc/movie-narrator.git
cd movie-narrator
pip install -e .

# Scene detection (PySceneDetect)
pip install "movie-narrator[media]"

# WhisperX + faster-whisper + FunASR + semantic search (requires PyTorch; Python < 3.14)
pip install "movie-narrator[ml]"

# Web UI (FastAPI + React) — separate package
pip install movie-narrator-web

# Everything
pip install "movie-narrator[full]"

Note on Python 3.14+: The [ml] extra is limited to Python < 3.14 — PyTorch itself is now 3.14-ready (2.10+), but downstream ML dependencies (WhisperX, FunASR) have not yet shipped Python 3.14 wheels; on 3.14+ it is silently skipped and the align/match steps automatically soft-degrade (see Soft steps).

For development:

pip install -e ".[dev]"

• LLM: Default uses local Ollama (ollama serve to start). Or configure remote LLM via .env file.

# Generate a narrated movie video
mn create --movie "飞驰人生" --style "热血搞笑" --duration 60

# With custom voice and format
mn create --movie "飞驰人生" --voice "zh-CN-XiaoxiaoNeural" --format "9:16"

mn create --config examples/job.example.yaml     # Drive from YAML config
mn create --subtitle-lang en --subtitle-mode bilingual  # Multi-language subtitles
mn race --movie "飞驰人生" --video movie.mp4 --candidates 3  # Multi-candidate horse race
mn imitate --reference viral_ref.mp4 --movie "飞驰人生"  # Reference video imitation
mn serve               # Start remote inference API server (v0.6.1+)
mn submit -m <movie>   # Submit async task
mn tasks               # List recent tasks
mn version             # Show version
mn --help              # Show full CLI help

All CLI flags and usage examples are documented in examples/cli-usage.sh.

All settings use the MN_ prefix to avoid conflicts with other tools.

~/.movie-narrator/.env is auto-created with default values on first run — edit it to configure LLM, TTS, and other settings. This file lives outside the package, so pip install/upgrade/uninstall never touches it. You can also create a project-level .env in your working directory for per-project overrides.

MN_LLM_BASE_URL=http://localhost:11434/v1
MN_LLM_API_KEY=ollama
MN_LLM_MODEL=qwen2.5:7b
MN_DEFAULT_VOICE=zh-CN-YunxiNeural

# PowerShell
$env:MN_LLM_BASE_URL="http://localhost:11434/v1"
$env:MN_LLM_MODEL="qwen2.5:7b"
mn create --movie "飞驰人生" --duration 60

# Linux / macOS
export MN_LLM_BASE_URL=http://localhost:11434/v1
export MN_LLM_MODEL=qwen2.5:7b
mn create --movie "飞驰人生" --duration 60

See .env.example for the complete list of all environment variables (LLM + TTS infrastructure only). All pipeline behavior is configured via examples/job.example.yaml — params keys covering scene detection, match, render, translate, BGM, WhisperX/FunASR align, async, and video sizes.

Movie Narrator works with any OpenAI-compatible LLM. New user? Check out the LLM Provider Guides for step-by-step registration and free-tier setup:

clips/ holds a standalone clip per segment, ready for secondary editing or reuse.

16-step sequential pipeline (see Architecture):

resolve_video → prepare_assets → research_plot → generate_script →
export_script_md → generate_voice → align_audio → detect_scenes →
match_clips → mix_bgm → translate_subtitles → generate_subtitle →
run_qa_gate → render_video → validate_deliverable → export_clips

Soft steps (research, align, scene detect, scene match, BGM, translate, QA gate, clip export) gracefully skip or soft-degrade when optional dependencies are missing or upstream data is unavailable. Use --strict to abort instead.

For contributors — most users only need mn create. See Architecture for details.

movie-narrator/
├── src/movie_narrator/
│   ├── cli.py               # Typer CLI entry point
│   ├── config.py            # Pydantic settings
│   ├── models.py            # Data models (Context, Status, etc.)
│   ├── contract.py          # Stable API contract surface
│   ├── pipeline/            # 16-step pipeline (runner, steps, errors)
│   ├── cloud/               # Task queue, remote inference, batch, scheduling, DLQ, distributed (v0.9.x)
│   ├── reliability/         # Circuit breaker + retry policy (v0.9.1)
│   ├── workflow/            # YAML job config (schema, loader, merge)
│   ├── tts/                 # TTS provider abstraction layer
│   └── utils/               # Shared utilities (console, log, font, etc.)
├── tests/                   # Unit + integration tests
├── docs/                    # Architecture, guides, roadmap
├── examples/                # Job YAML, CLI usage, plugins
└── .github/workflows/       # CI/CD

• Roadmap
• Architecture
• LLM Provider Guides
• Contributing
• AI Coding Assistant Guide

• Edge-TTS: the default TTS channel (edge) is built on a reverse-engineered, unofficial interface. It is provided for personal / non-commercial free testing only. For commercial deployment, switch MN_TTS_PROVIDER to openai or mimo (both built-in).
• TMDB: movie research data is sourced from TMDB (The Movie Database). When TMDB data is used, the source attribution is recorded in research.json as required, as a courtesy.

Licensed under the AGPL-3.0-or-later License.