MeshCoder: LLM-Powered Structured Mesh Code Generation from Point Clouds

Reconstructing 3D objects into editable programs is pivotal for applicationslike reverse engineering and shape editing. However, existing methods oftenrely on limited domain-specific languages (DSLs) and small-scale datasets,restricting their ability to model complex geometries and structures. Toaddress these challenges, we introduce MeshCoder, a novel framework thatreconstructs complex 3D objects from point clouds into editable Blender Pythonscripts. We develop a comprehensive set of expressive Blender Python APIscapable of synthesizing intricate geometries. Leveraging these APIs, weconstruct a large-scale paired object-code dataset, where the code for eachobject is decomposed into distinct semantic parts. Subsequently, we train amultimodal large language model (LLM) that translates 3D point cloud intoexecutable Blender Python scripts. Our approach not only achieves superiorperformance in shape-to-code reconstruction tasks but also facilitatesintuitive geometric and topological editing through convenient codemodifications. Furthermore, our code-based representation enhances thereasoning capabilities of LLMs in 3D shape understanding tasks. Together, thesecontributions establish MeshCoder as a powerful and flexible solution forprogrammatic 3D shape reconstruction and understanding.