ETRI-Knowledge Sharing Plaform

KOREAN
논문 검색
Type SCI
Year ~ Keyword

Detail

Conference Paper ScriptStage: Compiling Theater Scripts into Synchronized Generative-AI Previsualizations
Cited - time in scopus Share share facebook twitter linkedin kakaostory
Authors
Sang-Woo Seo, Haedong Kim, Seokhee Jeon
Issue Date
2026-10
Citation
International Conference on Information and Communication Technology Convergence (ICTC) 2026, pp.1-4
Publisher
IEEE
Language
English
Type
Conference Paper
Abstract
We present ScriptStage, a compiler-style system that turns a theater script into a set of synchronized directing assets for previsualization—a rough, geometrically coherent preview of a scene made before full production. The contribution is the compilation architecture, not a new generative model: a large language model (LLM) turns the script into a typed, checkable directing plan; a deterministic checker verifies that the available generative models can actually produce that plan; and an alignment step places the independently generated clips into one shared coordinate frame. Across three productions, all seven directing elements (five generated, two selected from an asset library) are produced and placed on one shared timeline. In a pilot with three open-weight LLMs (31–235B) on synthetic and real theater scripts, the checker reliably catches plans the models cannot execute—more often on real scripts than on synthetic ones. Among the two automatically scored violation classes, all invalid first-pass plans violated zone continuity, identifying cross-segment actor placement as the dominant observed bottleneck in this pilot. A lightweight rule-targeted repair loop then fixes 17/30 persistent failures, usually in one retry. ScriptStage thus shows that checking each plan against what the generative models can do, then geometrically aligning the results, makes heterogeneous generative-AI models work as one synchronized previsualization.
Keyword
generative-AI orchestration, large language models, multimodal AI systems, capability validation, smart-venue applications, stage previsualization
KSP Keywords
Language Models, generative model