Reproduce “PivotMesh: Generic 3D Mesh Generation via Pivot Vertices Guidance” 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
Haohan Weng, Yikai Wang, Tong Zhang, C. L. Philip Chen, Jun Zhu
PivotMesh generates polygonal 3D meshes directly by first predicting a coarse set of degree-selected pivot vertices and then completing the full mesh. Its hierarchical Transformer tokenizer and pivot-guided generation improve reconstruction and distributional metrics over PolyGen and MeshGPT on ShapeNet, Objaverse, and Objaverse-XL, while also supporting reference-based variation and coarse-to-detailed mesh refinement.
Key insights
- Representing shared vertices explicitly helps preserve mesh continuity better than treating every face corner as an independent token; hierarchical face-to-vertex decoding further improves reconstruction.
- High-degree vertices offer a compact structural summary for coarse-to-fine generation, and removing pivot guidance or replacing degree-based selection with random pivots degrades coverage and matching quality.
- Pivot conditioning preserves broad shape structure while allowing multiple plausible local completions, making the method useful for mesh variation but limiting precise geometric control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.