SYN3D-LLM: Synthetic 3D Point Clouds Generated Through LLM Discussion

Abstract: Creating synthetic datasets is challenging as they need to represent the diversity of real-world sites and measured point clouds. This paper introduces SYN3D-LLM, a dialog-driven large language model framework for generating semantically annotated synthetic 3D point clouds. By producing executable code through an iterative designer-critic loop. Experimental evaluations of the framework demonstrate that it improves the pass rate by discussing and testing generated code and by assessing the geometrical consistency between the initial prompt and the generated data. It also demonstrates that SYN3D-LLM is versatile in generating point clouds for various cases.