Recently, 3D assets created through reconstruction and generation have reached the quality level of manually crafted assets, highlighting their potential in alternative fields. However, this potential has not been fully realized, as these assets always require conversion into meshes for use in the 3D industry, and the current mesh extraction methods produce meshes that are significantly inferior to those created by human artists (AMs). Specifically, current mesh extraction methods rely on dense faces and overlook geometric features, leading to inefficient, complex post-processing, and lower representation quality.
To address these issues, researchers have proposed MeshAnything, an autoregressive model for generating artist-created 3D meshes. MeshAnything seamlessly integrates with various existing models to generate high-quality text/image/shape-conditioned mesh generation.
Product Entry:https://top.aibase.com/tool/meshanything
The meshes generated by MeshAnything significantly improve storage, rendering, and simulation efficiency while achieving comparable accuracy to previous methods.
The architecture of MeshAnything includes a VQ-VAE and a shape-conditioned decoder-only transformer. First, a VQ-VAE learns the mesh vocabulary, and then a shape-conditioned decoder-only transformer is trained on this vocabulary for shape-conditioned autoregressive mesh generation. Extensive experiments demonstrate that this method generates AMs with hundreds of times fewer faces than previous methods, significantly improving storage, rendering, and simulation efficiency, while achieving comparable accuracy to previous methods.
By integrating with various 3D asset production methods, MeshAnything enables highly controllable artist-created mesh generation. Additionally, compared to ground truth, this method has advantages in mesh topology and face count, and can generate meshes with completely different topologies but similar shapes, proving that the method does not merely overfit but understands how to construct meshes with efficient topologies.
Core features of this product include:
Powerful Mesh Generation: MeshAnything leverages autoregressive transformer technology to convert various inputs, such as images and point clouds, into fine-grained mesh models, with outstanding generation capabilities and model representation.
Automated Art Creation: MeshAnything provides users with convenient tools, making art creation more automated and intelligent, allowing users to focus on creative expression without being overly concerned with technical details.
Versatile Applications: MeshAnything has a wide range of applications in various fields, including industrial design, art creation, digital entertainment, and more, meeting the creative and needs of different users.
It should be noted that MeshAnything requires approximately 7GB and 30 seconds to generate meshes on an A6000 GPU. Limited by computational resources, MeshAnything is only trained on meshes with fewer than 800 faces and cannot generate meshes with more than 800 faces. The shape of the input mesh must be sufficiently clear, otherwise, it will be very difficult to represent it with only 800 faces. Therefore, feedforward image-to-3D methods often produce poor results due to insufficient shape quality.
Try it out: https://huggingface.co/spaces/Yiwen-ntu/MeshAnything
