CraftsMan3D is a novel 3D generative system that produces high-fidelity 3D geometries with varied shapes, detailed surfaces, and interactive refinement capabilities. Inspired by the workflow of a craftsman, the system first generates a coarse 3D shape using a 3D-native DiT model, then enhances surface details through either an automatic or interactive geometry refiner. The method addresses challenges in 3D generation, such as long optimization times, self-occlusion, and difficulties in accommodating user editing. A robust data preprocessing pipeline, integrating visibility checks and winding number analysis, maximizes the use of existing 3D data. The system leverages this data to train a 3D-native DiT model, which generates coarse geometries quickly. A normal-based geometry refiner then enhances local surface details, either automatically or interactively. Extensive experiments show that CraftsMan3D produces high-quality 3D meshes more efficiently than existing methods. The method also includes a novel normal-based interactive mesh refiner that can enhance geometries within 20 seconds, supporting user interaction to refine the generated 3D models. The system is evaluated on various datasets, demonstrating superior performance in mesh generation and refinement, particularly in handling complex geometries and self-occluded regions. The method's efficiency and effectiveness make it a promising approach for 3D modeling applications.CraftsMan3D is a novel 3D generative system that produces high-fidelity 3D geometries with varied shapes, detailed surfaces, and interactive refinement capabilities. Inspired by the workflow of a craftsman, the system first generates a coarse 3D shape using a 3D-native DiT model, then enhances surface details through either an automatic or interactive geometry refiner. The method addresses challenges in 3D generation, such as long optimization times, self-occlusion, and difficulties in accommodating user editing. A robust data preprocessing pipeline, integrating visibility checks and winding number analysis, maximizes the use of existing 3D data. The system leverages this data to train a 3D-native DiT model, which generates coarse geometries quickly. A normal-based geometry refiner then enhances local surface details, either automatically or interactively. Extensive experiments show that CraftsMan3D produces high-quality 3D meshes more efficiently than existing methods. The method also includes a novel normal-based interactive mesh refiner that can enhance geometries within 20 seconds, supporting user interaction to refine the generated 3D models. The system is evaluated on various datasets, demonstrating superior performance in mesh generation and refinement, particularly in handling complex geometries and self-occluded regions. The method's efficiency and effectiveness make it a promising approach for 3D modeling applications.