Title: ShapeLib: Leveraging Large Language Models to Construct Programmatic 3D Shape Abstraction Libraries

Abstract:

This paper introduces ShapeLib, pioneering the use of Large Language Model (LLM) priors to engineer libraries of programmatic 3D shape abstractions. The system is designed to accommodate two types of user-defined design intents: a foundational set of exemplar shapes and high-level textual descriptions outlining the desired functions for the output library. Through a structured LLM workflow, ShapeLib identifies an abstraction library that aligns with these intents. This process begins with the LLM proposing various methods for applying and implementing functions, followed by a validation phase to ensure these functions effectively represent the seed shapes.

To generalize beyond the initial seed set, we have developed recognition networks specific to each library. These networks map input shapes—whether represented as primitives, voxels, or point clouds—to programs that utilize the newly identified abstractions. Our findings across diverse modeling domains, categorized by shape type, indicate that LLMs, when integrated with geometric reasoning, can be effectively guided to create abstraction function libraries that generalize well across different shape distributions.

ShapeLib advances the longstanding goal of shape analysis: the discovery of reusable, programmatic abstractions that offer interpretable and semantically consistent interfaces. Comprehensive evaluations highlight that ShapeLib outperforms previous abstraction discovery methods in terms of plausibility during manipulation, usability, and generalization capabilities. Furthermore, we showcase how ShapeLib’s abstraction functions facilitate various downstream applications, enabling shape editing and generation workflows by combining LLM-based reasoning on shape programs with geometry processing tools.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC