Research map / 3D vision / Face and shape
PivotMesh: Generic 3D Mesh Generation via Pivot Vertices Guidance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PivotMesh, a transformer-based hierarchical mesh auto-encoder plus an autoregressive generator that first produces degree-selected pivot vertices and then the full mesh tokens, scaling native mesh generation to Objaverse and Objaverse-XL.
Problem
Native mesh generation (modeling triangle sequences directly) has been limited to small, narrow-category datasets such as ShapeNet because mesh topology is complex and varied; the paper aims to scale it to large, diverse 3D datasets.
Approach
Replace GNN/CNN tokenizers with a Transformer auto-encoder that decodes faces then vertices hierarchically, and factor generation coarse-to-fine by first generating high-degree pivot vertices that condition the full mesh token sequence in one autoregressive Transformer.
Key results
- On ShapeNet mixed set, PivotMesh reaches 50.42% COV, 11.03e-3 MMD, 60.89% 1-NNA versus MeshGPT's 43.89%, 11.48e-3, 63.24%.
- On Objaverse, PivotMesh reaches 46.48% COV and 58.55% 1-NNA versus MeshGPT's 35.03% and 63.86%.
- Auto-encoder triangle accuracy is 97.89% with L2 distance 1.05, versus 92.30% and 5.37 without hierarchical decoding.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PivotMesh instead of MeshGPT | unconditional mesh generation · ShapeNet Mixed | COV (%): 43.89 -> 50.42 | Helped |
| use PivotMesh instead of MeshGPT | unconditional mesh generation · ShapeNet Mixed | MMD (10^-3): 11.48 -> 11.03 | Helped |
| use PivotMesh instead of MeshGPT | unconditional mesh generation · ShapeNet Mixed | 1-NNA (%): 63.24 -> 60.89 | Helped |
| use PivotMesh instead of MeshGPT | unconditional mesh generation · Objaverse | COV (%): 35.03 -> 46.48 | Helped |
| use PivotMesh instead of MeshGPT | unconditional mesh generation · Objaverse | MMD (10^-3): 17.30 -> 16.66 | Helped |
| use PivotMesh instead of MeshGPT | unconditional mesh generation · Objaverse | 1-NNA (%): 63.86 -> 58.55 | Helped |
| use PivotMesh instead of MeshGPT | unconditional mesh generation · Objaverse-xl | COV (%): 41.50 -> 45.30 | Helped |
| use PivotMesh instead of MeshGPT | unconditional mesh generation · Objaverse-xl | MMD (10^-3): 14.76 -> 14.33 | Helped |
| use PivotMesh instead of MeshGPT | unconditional mesh generation · Objaverse-xl | 1-NNA (%): 64.25 -> 57.75 | Helped |
| use degree-based pivot selection instead of random pivot selection (ablation) | PivotMesh · Objaverse | COV (%): 42.48 -> 46.48 | Helped |
| use degree-based pivot selection instead of random pivot selection (ablation) | PivotMesh · Objaverse | MMD (10^-3): 17.62 -> 16.66 | Helped |
| use degree-based pivot selection instead of random pivot selection (ablation) | PivotMesh · Objaverse | 1-NNA (%): 62.90 -> 58.55 | Helped |
Surprising findings
- Randomly selected pivot vertices can perform worse than having no pivot guidance on some metrics (MMD and 1-NNA), per the ablation.
Limitations
- Controllability is limited because the same pivot vertices can correspond to many different meshes, producing undesired geometry.
- Training used only meshes under 500 faces on Objaverse and Objaverse-xl and a model much smaller than LLMs due to compute constraints.
Experimental setup
Datasets: ShapeNetV2 Objaverse Objaverse-xl
Benchmarks: ShapeNet Objaverse Objaverse-xl
Metrics: COV MMD 1-NNA triangle accuracy L2 distance
Baselines: PolyGen MeshGPT InstantMesh
Builds on: MeshGPT PolyGen residual vector quantization
Closest papers by meaning
- MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers (CVPR 2024 · similarity 0.89)
- ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.89)
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.87)
- MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers (ICLR 2025 · similarity 0.87)
- MeshFormer : High-Quality Mesh Generation with 3D-Guided Reconstruction Model (NeurIPS 2024 · similarity 0.87)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.