Title: MeshFlow: Streamlining Artistic Mesh Creation through MeshVAE and Flow-Driven Diffusion Transformers

Abstract:

This paper introduces MeshFlow, a novel approach designed to produce high-quality, artist-style 3D meshes. Traditional mesh generation techniques frequently rely on Auto-Regressive (AR) next-token prediction, a strategy that aligns well with the discrete characteristics of mesh topology. However, AR models suffer from significant scalability limitations, as their inference costs grow quadratically with the size of the mesh. Additionally, these methods necessitate the discretization of vertex coordinates, a process that inevitably leads to quantization errors.

To overcome these obstacles, we propose a Variational Autoencoder (VAE) supervised by a contrastive loss function. This architecture maps both the continuous positions of vertices and their discrete connectivity structures into a unified, continuous latent space. This resulting latent representation is substantially more compact than existing token-based methods. Leveraging this efficient representation, we developed a 3D generator powered by a Rectified Flow transformer. Unlike sequential approaches, our model synthesizes all mesh vertices and edges simultaneously in parallel. Benchmarks indicate that MeshFlow operates 18 times faster than the leading AR generator, while also delivering superior performance across standard metrics for mesh generation.

Homepage: https://mesh-flow.github.io/ Code: https://github.com/facebookresearch/meshflow

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC