SHOW-1 and Showrunner Agents in Multi-Agent Simulations

This approach generates full episodic content for existing intellectual properties by combining large language models, custom diffusion models and a multi-agent simulation that handles context, story progression and behavioral control. Showing users only a finished episode, rather than requiring repeated accept-or-reject choices, is presented as a way to avoid the "slot machine effect" of instant generation tools, which can cause frustration and loss of control. The system uses GPT-4 for agents and scene generation through prompt chains, and two DreamBooth-trained diffusion models were built from about 1,200 South Park characters and 600 background images.