CraftsMan3D is a novel 3D generative system that produces high-fidelity 3D geometries with varied shapes, detailed surfaces, and interactive refinement capabilities. The system is inspired by the workflow of a craftsman, who first creates a rough shape and then refines the surface details. It uses a 3D-native DiT model to generate coarse geometries from single images or text prompts, followed by an interactive geometry refiner that enhances surface details. The system addresses challenges in 3D generation, such as long optimization processes, self-occlusion, and irregular mesh topologies, by integrating visibility checks and winding number analysis to improve the success rate of watertight mesh conversion. The system also employs a normal-based geometry refiner that allows for interactive manipulation, enhancing the generated geometries to better align with user intentions. The method achieves high efficacy in producing high-quality 3D meshes compared to existing methods, with results validated through extensive experiments and visual comparisons. The system is efficient, with a generation time of just 25 seconds, and supports interactive refinement, making it a promising solution 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. The system is inspired by the workflow of a craftsman, who first creates a rough shape and then refines the surface details. It uses a 3D-native DiT model to generate coarse geometries from single images or text prompts, followed by an interactive geometry refiner that enhances surface details. The system addresses challenges in 3D generation, such as long optimization processes, self-occlusion, and irregular mesh topologies, by integrating visibility checks and winding number analysis to improve the success rate of watertight mesh conversion. The system also employs a normal-based geometry refiner that allows for interactive manipulation, enhancing the generated geometries to better align with user intentions. The method achieves high efficacy in producing high-quality 3D meshes compared to existing methods, with results validated through extensive experiments and visual comparisons. The system is efficient, with a generation time of just 25 seconds, and supports interactive refinement, making it a promising solution for 3D modeling applications.