PrintAnything: Learning Geometric Plan Map for 3D Printing G-code Generation from Unoriented Point Clouds

Traditional 3D printing pipelines require watertight meshes as input. Because point clouds—the native output of many sensors—are not watertight and often contain noise, they must first be converted into meshes. This reconstruction process is frequently ill-posed, leading to topological errors or geometric artifacts that cause printing failures. This paper asks: can we generate executable G-code directly from raw, unoriented point clouds without the intermediate step of mesh reconstruction?

The authors propose PrintAnything, a framework that treats G-code generation as a learning task. The core innovation is the Geometric plan (G-plan) map, a compact 2D representation for each slice of a 3D object. This map encodes three critical components: an occupancy map (where material is printed), a region map (structural categories like walls or infill), and a flow map (extrusion properties).

To bridge the gap between 3D point clouds and 2D slice-wise printing, the authors introduce a slice-wise point projection strategy. This method projects points near a specific slice height onto a 2D grid, providing the model with local geometric context. A neural network, GPNet, predicts the G-plan maps for each slice. Finally, the system uses a lightweight recommender to select optimal infill patterns and compiles the maps into standard G-code instructions.

PrintAnything outperforms traditional "point cloud-to-mesh-to-slicer" pipelines on the Slice-100K dataset. By eliminating the mesh reconstruction step, the model avoids the propagation of geometric artifacts that typically lead to printing failures. Quantitative results show superior performance in both 3D geometric accuracy (Chamfer Distance) and 2D slice-level structural fidelity. The resulting G-code is directly executable by standard 3D printers, providing a more robust and practical workflow for fabrication from raw sensor data.