Enhancing movie script creation through retrieval-augmented LLMs and stable diffusion scene modeling.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Script writing is a complex, labor-intensive process requiring creativity and attention to detail.
• Advancements in Natural Language Processing (NLP) and Deep Learning offer potential for automating script generation.
• Existing methods may struggle with context-specific details and creative customization.

Purpose of the Study:

• To explore automated movie script and scene generation using advanced AI techniques.
• To evaluate the effectiveness of Retrieval Augmented Generation (RAG) and LLM fine-tuning for scriptwriting.
• To assess the coherence, relevance, and quality of AI-generated scripts and scenes.

Main Methods:

• Implemented Retrieval Augmented Generation (RAG) with Gemini-Pro, storing scripts in a vector database for context.
• Employed fine-tuning on Large Language Models (LLMs) like GPT-2 and Bloom for script generation.
• Integrated Stable Diffusion and CompVis for AI-driven scene generation based on script elements.

Main Results:

• Gemini-Pro (RAG) achieved a cosine similarity of 0.5713 between prompts and generated scripts.
• GPT-2 and Bloom (fine-tuning) showed cosine similarities of 0.5011 and 0.5058, respectively.
• GPT-2 demonstrated strong coherence and relevance (perplexity 1.7443), while CompVis achieved a CLIP score of 0.3061 for scene generation.

Conclusions:

• Retrieval Augmented Generation (RAG) with Gemini-Pro is effective for context-specific script generation.
• Fine-tuned LLMs like GPT-2 can produce coherent and relevant movie scripts.
• AI models show promise in automating scriptwriting and extending to scene generation, enhancing creative prototyping.