混合實境(MR)系統正逐步成為協同創作的夥伴,然而要將語音請求即時轉化為互動式的 3D 內容依然困難,因為現有系統多仰賴固定資產庫、僅能組合引擎基本物件,或只能為預先綁定骨骼的物件添加動畫。
本文提出一套基於遠端渲染架構的整合式 MR 系統,結合對話式大型語言模型(LLM)與多代理(multi-agent)後端,能夠在即時場景中從頭合成任意的 3D 物件,並賦予其互動能力。專門的代理將語音驅動的創作過程拆解為特定階段:3D Crafter 透過由視覺語言模型(VLM)審查員把關的視覺審查流程,將幾何形狀直接合成為可執行的程式碼;動畫工具(Animation Tool)則針對具語意命名的子部件組合基本動作,為動態生成的網格注入符合情境的動態效果。
在包含 50 個開放式提示詞的基準測試中,結構化的提示方法將首次生成的錯誤率降至 1\%;此外,一項涵蓋三個不同背景群體、共 14 名參與者的使用者研究顯示,該系統的整體體驗評分達到 4.07 分(滿分 5 分)。;Mixed Reality (MR) systems are beginning to act as co-creative partners, yet turning a spoken request into interactive 3D content in real time remains hard, because existing systems retrieve assets from fixed libraries, assemble engine primitives, or can only animate objects with predefined structures.
Specialized agents decompose voice-driven creation into bounded stages: a 3D Crafter synthesizes geometry directly as executable code gated by a vision-language model (VLM) visual reviewer, and an Animation Tool injects contextually appropriate motion into dynamically generated meshes by composing primitive actions over semantically named sub-parts.
On a benchmark of 50 open-ended prompts, structured prompting reduces the first-pass generation error rate to 1\%, and a user study of 14 participants across three background groups rates the overall experience at 4.07 out of 5.