MeshAnything is a model designed for generating Artist-Created Meshes (AMs) using autoregressive transformers.
MeshAnything addresses the limitations of current mesh extraction methods, which often produce meshes with dense faces and ignore geometric features, leading to inefficiencies and lower representation quality. MeshAnything treats mesh extraction as a generation problem, producing AMs that align with specified shapes and can be integrated with various 3D asset production methods, enhancing their application across the 3D industry.
Key Features and Architecture of MeshAnything
- Architecture: MeshAnything comprises a VQ-VAE (Vector Quantized Variational Autoencoder) and a shape-conditioned decoder-only transformer. The VQ-VAE learns a mesh vocabulary, which is then used by the shape-conditioned transformer for autoregressive mesh generation.
- Efficiency: The model generates meshes with hundreds of times fewer faces compared to traditional methods, significantly improving storage, rendering, and simulation efficiencies while maintaining precision comparable to previous methods.
- Training Data: MeshAnything is trained on a combined dataset from Objaverse and ShapeNet, filtered to include only meshes with fewer than 800 faces. This dataset includes 51,000 meshes from Objaverse and 5,000 from ShapeNet.
- Qualitative Performance: Experiments show that MeshAnything can generate meshes with efficient topology that conform to given shapes, often surpassing the quality of manually created meshes. The model does not simply overfit but understands how to construct meshes using efficient topology.
MeshAnything represents a significant advancement in the field of 3D asset generation, offering a scalable and efficient solution for creating high-quality meshes that can be easily integrated into various 3D production pipelines.