Researchers from Nanyang Technological University, Tsinghua University, Imperial College London, and Westlake University introduced MeshAnything V2, an AI model that significantly improves the generation of artist-created meshes (AMs) aligned with given 3D shapes. This advancement marks a substantial leap in both performance and efficiency for 3D asset production.
Read the full paper here
The key innovation behind MeshAnything V2 is the newly proposed Adjacent Mesh Tokenization (AMT) method. Unlike previous approaches that represent each face of a mesh with three vertices, AMT uses a single vertex whenever possible, resulting in a more compact and well-structured token sequence.
Experiments show that AMT reduces the token sequence length by approximately half on average, significantly decreasing computational load and memory usage. This efficiency gain allows MeshAnything V2 to generate meshes with up to 1600 faces, doubling the previous limit of 800.
The researchers conducted extensive tests to validate the effectiveness of AMT. Quantitative experiments demonstrated that MeshAnything V2 outperforms its predecessor in various metrics, including Chamfer Distance, Edge Chamfer Distance, and Normal Consistency.
MeshAnything V2 is designed to be integrated with various 3D asset production pipelines, enabling highly controllable AM generation. This capability has potential applications in industries such as gaming, movies, and virtual reality, where high-quality 3D meshes are in demand.
While the advancements are significant, the researchers acknowledge that further improvements in stability and accuracy are needed before the technology is ready for industrial applications.
The development of MeshAnything V2 builds upon recent progress in 3D generation techniques, including the application of transformers and diffusion models to 3D asset creation. This research contributes to the growing field of AI-assisted 3D content generation, which aims to streamline and enhance the traditionally time-consuming process of manual mesh creation.
As the technology continues to evolve, it holds the promise of revolutionising 3D asset production, potentially reducing the reliance on manual labor and accelerating the creation of complex 3D environments for various digital media applications.
Moreover, this technology is particularly beneficial for companies and individuals involved in creating 3D models, as it enhances accuracy and efficiency. For instance, 3DAiLY, a 3D model creation company, uses generative AI to produce ultra-realistic, production-ready assets. They are also pioneers in making these models accessible on their community platform, catering to both AAA and indie game developers.
