ShapeLib: Designing a library of programmatic 3D shape abstractions with Large Language Models

ShapeLib is a new method that leverages Large Language Models to automatically design libraries of reusable 3D shape abstractions from user-provided descriptions and exemplar shapes. The system validates these abstractions through geometric reasoning and develops recognition networks that generalize across shape distributions, enabling interpretable programmatic interfaces for 3D modeling tasks.

ShapeLib represents a significant advancement in computational geometry by bridging LLM capabilities with structured 3D shape analysis. The approach addresses a long-standing challenge in computer vision and graphics: discovering reusable, interpretable abstractions for complex shape families without manual engineering. Rather than treating LLMs as black boxes, the researchers thoughtfully integrate geometric reasoning to guide LLM-generated shape functions, ensuring they remain both semantically meaningful and practically useful.

This work emerges from growing recognition that LLMs encode valuable prior knowledge about how concepts decompose and relate hierarchically. The seed-driven design pattern—accepting both natural language descriptions and exemplar shapes—makes the system accessible to domain practitioners while maintaining mathematical rigor. The development of library-specific recognition networks demonstrates how to bridge from unstructured shape data (primitives, voxels, point clouds) to structured programmatic representations.

For practitioners in 3D modeling, CAD, computational design, and game development, ShapeLib's approach could significantly reduce the manual effort required to build shape abstraction systems. The framework's demonstrated advantages in generalization, usability, and plausibility under manipulation suggest practical deployment potential. The downstream applications combining LLM reasoning with geometry processing tools hint at future workflows where AI assists in shape editing and generation tasks.

The research positions a new frontier where AI systems don't replace human design expertise but rather scaffold and accelerate abstraction discovery. Future work likely involves scaling to more complex shape domains, integrating with commercial 3D software pipelines, and exploring how these discovered abstractions transfer across different modeling contexts and disciplines.

• →ShapeLib uses LLMs with geometric reasoning to automatically discover reusable 3D shape abstractions from text descriptions and example shapes.
• →The system validates discovered abstractions through library-specific recognition networks that generalize across different shape distributions.
• →Integration of LLM priors with structured geometric validation produces interpretable, semantically aligned programmatic interfaces for 3D modeling.
• →Framework demonstrates practical advantages in generalization and usability compared to prior abstraction discovery methods.
• →Enables downstream applications combining shape programming with geometry processing for editing and generation workflows.