MeshAnything V2: Generating Artist-Created Meshes with Fewer Tokens
đź“„ Full Paper
đź’¬ Ask
Artist-created meshes (AMs) are the standard 3D representations for many industries, from gaming to virtual reality. They provide a level of control and detail that is difficult to match with other methods. However, manually creating these meshes is a labor-intensive, time-consuming process. Researchers are increasingly turning to machine learning to automate the process, but existing methods have limitations.
The current methods for generating AMs suffer from significant inefficiencies and limitations. They treat a mesh as a sequence of faces, each of which contains three vertices. This makes for very large token sequences. Moreover, the existing methods have a hard limit on the number of faces they can generate.
A new research paper, “MeshAnything V2: Artist-Created Mesh Generation with Adjacent Mesh Tokenization,” presents a novel solution to these problems. It introduces a method called Adjacent Mesh Tokenization (AMT) that significantly reduces the size of the token sequences and increases the number of faces that can be generated.
AMT represents a face with a single vertex whenever possible. This is achieved by encoding the face and then encoding its adjacent face, which shares an edge. Only one additional vertex is needed to represent this adjacent face. If no adjacent face is found, AMT inserts a special token into the sequence to mark this event and restarts from a face that has not been encoded yet.
In effect, AMT can reduce the length of the token sequence by half on average, resulting in a substantial reduction in the computational load and memory usage.
The research team also introduces MeshAnything V2, a new model equipped with AMT. It significantly outperforms the previous version of MeshAnything. MeshAnything V2 can generate meshes up to 1,600 faces, double the previous limit. The model is also more efficient and generates high-quality, highly controllable AMs.
“Our experiments show that AMT can significantly improve both the efficiency and performance of mesh generation,” the researchers said. “The results are a major step forward in the development of automated mesh generation techniques. We believe that our work will contribute to the development of more efficient and high-quality 3D content creation tools.”
The researchers have made MeshAnything V2 open-source, which means that other researchers can use it to build upon their work and develop even more advanced techniques for generating artist-created meshes. This development has the potential to revolutionize the 3D content creation industry, making it easier and faster to create realistic and detailed 3D models for a wide range of applications.
Chat about this paper
To chat about this paper, you'll need a free Gemini API key from Google AI Studio.
Your API key will be stored securely in your browser's local storage.