Dramatron addresses the challenge of generating coherent scripts and screenplays by leveraging large language models (LLMs). It serves as a co-writing tool and source of inspiration for playwrights and screenwriters, aiding in creative exploration and idea generation.

How It Works

The system employs hierarchical story generation to maintain consistency across lengthy narrative outputs. It interactively generates elements like character descriptions, plot points, locations, and dialogue, starting from a logline. This approach facilitates co-writing by providing structured material for human authors to compile, edit, and rewrite, offering a novel way to integrate LLMs into creative writing workflows.

Quick Start & Requirements

Dramatron can be run via a Python Colab notebook (colab/dramatron.ipynb). The provided notebook is "unplugged," requiring users to integrate their own LLM by implementing init and sample functions. Specific LLM integration and associated computational requirements (e.g., GPU, memory) depend on the chosen LLM.

Highlighted Details

• Evaluated through user studies with 15 professional playwrights and screenwriters.
• Output has been used in a public theatre production ("Plays by Bots").
• Users found it valuable for "world building," exploring alternative story paths, and creative idea generation.
• Offers a structured, top-down approach to script generation.

Maintenance & Community

Contact is available via dramatron@deepmind.com. A guide for contributors is provided. The project is associated with research published in SIG CHI 2023 and arXiv.

Licensing & Compatibility

Software components are licensed under the Apache License, Version 2.0 (Apache 2.0), which is permissive for commercial use. Other materials are licensed under the Creative Commons Attribution 4.0 International License (CC-BY).

Limitations & Caveats

The project requires users to integrate their own LLM. Its top-down hierarchical structure may not suit all writing processes. LLM outputs may contain biases, offensive text, or plagiarized content, necessitating human review and potential filtering (e.g., via Perspective API). User feedback indicates outputs can be "formulaic" and less suitable for writing entire plays autonomously.