SceneSage

LLM-powered CLI tool to analyze subtitle files (.srt) and extract structured metadata for each scene.

Features

• Scene segmentation based on pauses ≥ 4s
• Local LLM-based extraction of:
  • Summary
  • Characters
  • Mood
  • Cultural references

Requirements

• Python 3.8+
• A compatible GPU (optional but recommended for faster inference)
• Hugging Face Transformers-compatible language model (e.g., HuggingFaceH4/zephyr-7b-beta)

Setup

pip install -r requirements.txt

Usage

python scenesage.py plan9.srt --model HuggingFaceH4/zephyr-7b-beta --output scenes.json

Arguments

• plan9.srt: The subtitle file you want to analyze (must be in .srt format).

• --model: Name or path of a locally available Hugging Face model compatible with transformers.

• --output: (Optional) Output file path for the results. Defaults to scenes.json.

Output:

• Segment the .srt file into scenes based on 4-second pauses.

• Use the LLM to extract structured metadata for each scene.

• Save the results in a JSON file like:

[
  {
    "start": "00:00:01,000",
    "end": "00:00:08,000",
    "transcript": "...",
    "summary": "...",
    "characters": [...],
    "mood": "...",
    "cultural_refs": [...]
  },
  ...
]

The file is saved to the path provided via --output, or scenes.json by default.