Reproduce “MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai and 4 more
The MESHANYTHING framework introduces a shape-conditioned autoregressive transformer to generate artist-created meshes (AMs) from diverse 3D inputs. This method produces meshes with efficient topology and high geometric fidelity, directly addressing the incompatibility of automated 3D outputs with industrial pipelines.
Key insights
- Explicitly conditioning mesh generation on a target 3D shape simplifies the learning task, allowing the model to focus its capacity on understanding and generating efficient, artist-quality topology.
- Incorporating deliberately corrupted ground truth shapes and a noise-resistant decoder during training is crucial for making the generative model robust to the imperfect geometry of automatically produced 3D assets.
- Autoregressive transformers can effectively learn to produce sparse, structured mesh topologies that mimic human-crafted assets when provided with appropriate shape conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.