MeshFlow generates 3D meshes 18× faster with equivariant flow matching
Researchers propose MeshFlow, an equivariant flow matching model that generates triangle meshes directly as unordered sets, matching autoregressive quality at 18× inference speed.
MeshFlow is a mesh generation model from a Stanford and Cornell team that produces 3D triangle meshes directly, avoiding the autoregressive token sequences that slow existing generators. Posted to arXiv on June 23, the paper introduces an equivariant optimal-transport flow matching architecture that respects the symmetries inherent to triangle soups—arbitrary permutations of faces and vertices within each face—while achieving quality on par with state-of-the-art autoregressive methods.
Meshes remain one of the most common 3D scene representations in graphics pipelines, game engines, and simulation environments, but generating them with neural networks has historically required serializing the geometry into long token sequences. Autoregressive models then predict one token at a time, which scales poorly as mesh complexity grows. MeshFlow sidesteps this bottleneck by treating meshes as unordered sets of triangles and learning to generate them in parallel.
The architecture modifies the Diffusion Transformer to model a velocity field that maintains permutation equivariance. Because triangle soups have no canonical ordering of faces or vertices, any model that respects the underlying geometry must produce identical outputs under arbitrary reorderings. The authors introduce an optimal-transport-based training objective that eliminates supervision signals violating these symmetries, improving convergence and stability during training. MeshFlow delivers comparable mesh quality to leading autoregressive generators while running roughly 18 times faster at inference—a speedup that comes from generating all triangles in a single forward pass rather than sequential token prediction. For practitioners working with 3D content pipelines—whether in game asset generation, architectural visualization, or robotics simulation—the inference time reduction could meaningfully compress iteration cycles. The project page and code are available at https://qiisun.github.io/MeshFlow/.
