Project Report: Chain of Consciousness: Collaborative Storytelling and Image Generation

1. Inspiration

We wanted to create a collective storytelling experience that captures how a single line or choice can ripple throughout an entire narrative. The idea was to blend collaborative input, LLM-based text generation, and AI-powered artwork to visualize how stories can evolve in unexpected directions. Our main inspirations were:

• “Exquisite Corpse”–style creative writing games, where multiple participants build a story line by line.
  
• The chain-of-events concept, illustrating how each small narrative step can influence the entire story arc.
  
• The emerging possibilities of generative AI (both text and image) and how they can enhance creativity in a community setting.

2. What We Learned

1. Prompt Engineering

   • We discovered the importance of clear, focused prompts for Large Language Models (LLMs), ensuring they return valid JSON or structured outputs.
     
   • Handling edge cases where the LLM might produce code fences or unexpected text required thoughtful post-processing.

2. Image Generation & API Handling

   • Experimented with DALL·E 3 and learned how to craft effective image prompts—merging story metadata (genre, tone, style) with focal characters and settings.
     
   • Explored how to save images locally from an AI image API response.

3. Designing an Evaluation & Ranking System

   • Implemented a method to score stories on dimensions like plot cohesion, creativity, characters, etc., mapping a final numeric total to ranks (E → SS).

3. How We Built the Project

1. Architecture Overview

   • Discord/CLI: Users provide story lines or prompt the AI to generate them.
     
   • LLM Integration (OpenAI GPT): For generating next lines, extracting metadata (new characters, settings), and summarizing stories.
     
   • In-Memory Data Storage: A global dictionary that stores each story by storyId. Each story object includes metadata, lines, characters, settings, etc.
     
   • Finalization & DB Hook: Once a story finishes, we “finalize” it—applying a final AI-driven metadata pass—then remove it from memory (with an option to save in a database).

2. Workflow

   • create_story: Initializes the JSON structure and requests initial metadata from the LLM.
     
   • add_new_line_and_update: Appends new lines (user or AI), triggers character/setting extraction, updates summaries.
     
   • finalize_story: Uses the LLM to propose final metadata (title, genre, etc.), then moves the story out of memory.
     

3. Image Generation

   • A build_dalle_prompt function dynamically constructs a DALL·E 3 prompt from the story’s metadata, characters, and setting.
     
   • We then download and save the resulting image to disk using Python’s requests.

4. Challenges Faced

1. Ensuring Valid JSON from the LLM

   • The model sometimes returned code blocks or extra commentary. We overcame this by:
     
     • Prompting: “Respond only in valid JSON.”
       
     • Regex post-processing to strip code fences if needed.

2. Maintaining Story Consistency

   • Sometimes the AI would introduce contradictory details (like reusing character names differently).
     
   • We mitigated this by storing structured data (e.g., characters array) and letting the LLM refer back to the existing “story so far.”

3. AI Prompts for Images

   • Generating cohesive visuals that truly match the story required iterative experimentation with DALL·E prompts—balancing detail with creative freedom.
     
   • We learned that specifying tone (dark, whimsical), color palettes, and key visuals improved the results.

Overall, this project showed us how to blend user creativity with AI assistance—from line-by-line generation to a final illustrated product. We hope to expand it further, possibly adding more sophisticated editing tools or branching narratives in the future!

Built With

• dall-e
• discord
• networkx
• openai
• python
• sqlite