Abstract:  This research addresses high-resolution, part-level artistic mesh generation. While current autoregressive methods yield high-quality topology, their scalability to high resolutions is bottlenecked by the quadratic complexity of attention mechanisms. To resolve this, MeshPack, a part-level autoregressive framework, is proposed. By decomposing meshes into multiple sequences for iterative generation, this approach signifi-cantly reduces computational complexity and naturally boosts output resolution. Furthermore, an infor-mation passing module is introduced to fuse previously generated parts with the current one, ensuring global context awareness and cross-part consistency. For model training, FurniSet3D—a high-quality fur-niture dataset featuring refined geometry, clean topology, and semantic decomposition is constructed. Qualitative evaluations on FurniSet3D demonstrate that MeshPack generates fine-resolution meshes with superior topological quality and reasonable part decomposition, substantially outperforming existing methods in furniture generation.