MeshRipple: Structured Autoregressive Generation of Artist-Meshes

Junkai Lin^*, Hang Long^*, Huipeng Guo, Jielei Zhang, JiaYi Yang, Tianle Guo, Yang Yang, Jianwen Li, Wenxiao Zhang, Matthias Niessner, Wei Yang^†

^* Equal contribution. ^† Corresponding author.

CVPR 2026 Highlight

Generates artist-style 3D meshes autoregressively while keeping long-range surface structure intact — earlier approaches serialized mesh faces into sequences and processed them in a sliding window, which broke geometric consistency across distant parts of the mesh. MeshRipple instead uses a topology-aligned tokenization, a face-expansion strategy that keeps the growing surface connected, and sparse-attention memory to preserve long-range dependencies, producing meshes with better surface quality and more complete topology.

LATO.2: Factorized 3D Mesh Generation with Vertex and Topology Flow

Hang Long^*, Tianhao Zhao^*, Junkai Lin, Youjia Zhang, Huipeng Guo, Rendong Liang, Jiale Xu, Jozef Hladký, Matthias Nießner, Yuanming Hu, Wei Yang^†

^* Equal contribution. ^† Corresponding author.

Generates a 3D mesh in two stages: first it places the vertices, then it works out how to connect them into faces, conditioned on a shared voxel scaffold. Splitting the problem this way lets it generate higher-resolution meshes part-by-part and support topology-adaptive editing, where dragging a vertex automatically updates the surrounding connectivity — outperforming prior topology-aware mesh generation methods.