MeshAnything V2: Artist-Created Mesh Generation With Adjacent Mesh Tokenization

Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, Guosheng Lin

Abstract

MeshAnything V2 introduces Adjacent Mesh Tokenization (AMT), a novel representation method that significantly improves the efficiency of autoregressive 3D mesh generation. By exploiting adjacency in manifold meshes, AMT reduces sequence lengths by approximately 50%, allowing Large Language Model (LLM) architectures to generate complex, artist-quality meshes with up to 1,600 faces. This approach effectively doubles the geometric complexity achievable by previous state-of-the-art models without increasing computational overhead.

Introduction: The Complexity Ceiling in 3D Mesh Generation

In modern 3D pipelines, meshes are the preferred representation due to their efficiency in rendering, storage, and animation. However, generating "Artist-Created Meshes" (AMs)—characterized by clean, low-polygon topology—is a significant challenge for automated systems. Recent research has attempted to treat mesh generation as a sequence modeling task using Transformers, but these models have hit a "complexity ceiling."

Standard tokenization methods represent each triangular face with three independent vertices, and each vertex requires three coordinate tokens. This results in a sequence length nine times the number of faces. For a transformer, these long, redundant sequences are computationally expensive to process and difficult to learn, effectively limiting generated meshes to a very low polygon count, typically under 800 faces. This redundancy acts as noise, obscuring the underlying structural patterns of human-authored topology.

MeshAnything V2 transforms dense, unoptimized meshes into clean, artist-quality topologies, as seen in this robot model.

Methodology: Adjacent Mesh Tokenization (AMT)

The core innovation of MeshAnything V2 is Adjacent Mesh Tokenization (AMT). Instead of naively listing every vertex for every face, AMT exploits the fact that most faces in a manifold mesh share edges with their neighbors. Unlike text, which is inherently linear, meshes are graph-based; AMT effectively "flattens" the 3D graph into a 1D stream by following the surface of the mesh.

The AMT Algorithm

The model first sorts vertices and faces to create a predictable pattern. When generating a sequence, if a new face is adjacent to the previous one, the model only needs to provide the one new vertex required to complete the triangle. This reduces the per-face token requirement from nine to as few as three.

To handle complex topologies where a continuous "strip" of faces isn't possible, the authors introduce special tokens:

  • Break Token (&): Used when the model cannot find an adjacent face, allowing it to restart with a full three-vertex face.
  • Swap Token ($): Used to handle different edge-sharing configurations between adjacent faces.

Controllable Density

The model, based on a decoder-only transformer (OPT-350M), includes a "Face Count Condition." This allows users to specify an approximate target for the number of faces, giving artists direct control over the level of detail (LOD) and ensuring the output meets specific project requirements.

Key Results and Versatility

MeshAnything V2 demonstrates significant improvements across both quantitative and qualitative benchmarks. AMT achieves an average sequence length compression ratio of 0.497 compared to the baseline. This efficiency allows the model to scale to 1,600 faces without changing the model's parameter count or increasing the computational cost per face.

The model demonstrates high versatility, generating clean meshes from diverse inputs including point clouds, NeRFs, 3D Gaussian Splats, and text prompts.

Quantitatively, the model achieves a lower Chamfer Distance (0.874 vs 0.895) and higher Normal Consistency (0.933 vs 0.924) than versions without AMT. Qualitatively, the model produces meshes with significantly improved topology. In a blind user study, 67.8% of participants preferred the meshes generated by V2 over the original MeshAnything. Furthermore, the model serves as a versatile "meshifier," capable of converting various 3D representations—such as Point Clouds, NeRFs, and 3D Gaussian Splats—into production-ready low-poly meshes.

Conclusion and Future Work

MeshAnything V2 represents a significant step toward automating the creation of high-quality 3D assets. By optimizing the underlying representation through AMT, the researchers have shown that sequence efficiency is a critical factor in scaling the complexity of generated 3D geometry.

Despite this progress, the authors acknowledge that a "complexity ceiling" still exists; many production-grade assets require even higher polygon counts than 1,600. Future work will focus on further scaling the face count and refining the "artistic" feel of the generated topology to better match the specific stylistic choices made by human modelers in professional environments.

Summary updated

Abstract

We introduce MeshAnything V2, an autoregressive transformer that generates Artist-Created Meshes (AM) aligned to given shapes. It can be integrated with various 3D asset production pipelines to achieve high-quality, highly controllable AM generation. MeshAnything V2 surpasses previous methods in both efficiency and performance using models of the same size. These improvements are due to our newly proposed mesh tokenization method: Adjacent Mesh Tokenization (AMT). Different from previous methods that represent each face with three vertices, AMT uses a single vertex whenever possible. Compared to previous methods, AMT requires about half the token sequence length to represent the same mesh in average. Furthermore, the token sequences from AMT are more compact and well-structured, fundamentally benefiting AM generation. Our extensive experiments show that AMT significantly improves the efficiency and performance of AM generation. Project Page: https://buaacyw.github.io/meshanything-v2/
BibTeX
@inproceedings{chen2025meshanything,
  author = {Yiwen Chen and Yikai Wang and Yihao Luo and Zhengyi Wang and Zilong Chen and Jun Zhu and Chi Zhang and Guosheng Lin},
  title = {{MeshAnything V2: Artist-Created Mesh Generation With Adjacent Mesh Tokenization}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure