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

MeshAnything V2, a significant update to the original MeshAnything, introduces a novel tokenization method called Adjacent Mesh Tokenization (AMT). This advancement not only boosts performance and efficiency but also enhances the quality of artist-created meshes (AM) generated by the model. For practitioners working with 3D asset production pipelines, this means more control over mesh generation while maintaining high fidelity to given shapes.

Technical Overview

MeshAnything V2 is an autoregressive transformer designed to generate AMs aligned with specified shapes. The key innovation lies in its tokenization method:

• Adjacent Mesh Tokenization (AMT): Unlike traditional methods that use three vertices to represent a face, AMT uses a single vertex whenever possible. If this isn't feasible, it adds a special token & and restarts the sequence.
• Compact Representation: On average, AMT reduces the token sequence length by half compared to previous methods. This compact representation enhances sequence learning, leading to more efficient and higher-quality mesh generation.

Key Benefits

• Performance Improvement: MeshAnything V2 outperforms its predecessor in both efficiency and performance while using models of the same size.
• High-Quality Output: It can generate AMs with up to 1600 faces that are well-aligned with given shapes.
• Controllability: The model integrates seamlessly with various 3D asset production pipelines, providing high controllability over the generated meshes.

Implementation Details

• Model Architecture:

  • Transformer-Based: MeshAnything V2 leverages a transformer architecture to generate mesh sequences.
  • Tokenization: AMT is applied during the tokenization phase to create more compact and structured token sequences.
  • Training Data: The model is trained on a diverse dataset of artist-created meshes, ensuring it can handle a wide range of shapes and styles.

• Benchmarks:

  • Sequence Length Reduction: Experiments show that AMT reduces the average token sequence length by 50%, significantly improving computational efficiency.
  • Mesh Quality: The generated meshes are of high quality, with better alignment to given shapes compared to previous methods.

Example Use Cases

• 3D Asset Production: Game developers and filmmakers can use MeshAnything V2 to generate high-quality meshes for characters, environments, and props.
• Artistic Creations: Artists can leverage the model to create intricate and detailed 3D models with greater control over the final output.
• Research Applications: Researchers in computer graphics and machine learning can explore new ways to enhance mesh generation using AMT.

Conclusion

MeshAnything V2 represents a significant step forward in artist-created mesh generation. By introducing Adjacent Mesh Tokenization, it not only improves performance and efficiency but also enhances the quality and controllability of generated meshes. For practitioners in 3D asset production, this model offers a powerful tool to streamline workflows and achieve high-quality results.