Xanadu: Generative Media Pipelines for Immersive Participatory Theater

Abstract

We present Xanadu, a full-scale participatory theater production that integrated generative AI into a live musical. Over two weeks, 500 audience members contributed sketches, movements, and sounds that were transformed in real-time --via pipelines combining vision-language models, diffusion models, and moderation-- into images, 3D objects, poetry, and choreography rendered within an extended reality environment. Audience inputs were framed as ritual "offerings" to the Muses, with performers guiding participation and AI serving as interpretive intermediary. A hybrid hosting architecture combined controllable, research-driven models with fast foundation models, enabling 30–60 second generations while preserving stylistic and narrative coherence. We discuss design strategies and trade-offs that supported large-scale, group-level human-AI collaboration. This work contributes practical insights into deploying generative AI in live performance and highlights opportunities for designing AI systems that facilitate collective rather than individual creativity.