Prompt Engineering for X3D Object Creation with LLMs

Large Language Models (LLMs) are a new class of knowledge embodied in a computer and trained on massive amounts of human text, image, and video examples. As the result of a user prompt, these LLMs can generate generally coherent responses in several kinds of media and languages. Can LLMs write X3D code? In this paper we explore the ability of several leading LLMs to generate valid and sensible code for interactive X3D scenes. We compare the prompt results from three different LLMs to examine the quality of the generated X3D. We setup an experimental framework that uses a within-subjects repeated-measures design to create X3D from text prompts. We vary our prompt strategies and give the LLMs increasingly challenging and increasingly detailed scene requests.We assess the quality of the resulting X3D scenes including geometry, appearances, animations, and interactions. Our results provide a comparison of different prompt strategies and their outcomes. Such results provide early probes into the limited epistemology and fluency of contemporary LLMs in composing multi-part, animate-able 3D objects.