Research map / 3D vision / Face and shape
ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction
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 ARMesh, generating triangle meshes auto-regressively coarse-to-fine by reversing a generalized (simplicial-complex) QSlim simplification into a vertex-split sequence and training a transformer to predict these refinement operations.
Problem
Autoregressive mesh generation typically builds meshes face by face in arbitrary lexicographic or traversal order, which ignores geometry and human coarse-to-fine perception and gives no intermediate quality control.
Approach
Generalize meshes to simplicial complexes, adapt QSlim into GSlim that allows topology changes and reduces any mesh to a single point, record the reversed simplification as progressive simplicial complex vertex splits, tokenize them compactly with topological labels and constrained decoding, and train an AR transformer to predict next-level-of-detail refinements.
Key results
- On ShapeNet bench, ARMesh at 10% of AR steps reaches COV 41.07 and MMD 1.67, versus EdgeRunner (9.43, 10.32) and BPT (15.15, 3.02); full-generation COV 56.29 is comparable to baselines.
- Uses about 2556 tokens per shape versus 5840 for EdgeRunner and 3235 for BPT.
- Full generation takes about 1 minute on average; 50% simplification still gives decent results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ARMesh vertex-based refinement tokenization instead of BPT triangle-based tokenization | Unconditional mesh generation · ShapeNet bench, 10% autoregressive steps | COV: 15.15 -> 41.07 | Helped |
| Use ARMesh vertex-based refinement tokenization instead of BPT triangle-based tokenization | Unconditional mesh generation · ShapeNet bench, 20% autoregressive steps | COV: 26.65 -> 47.40 | Helped |
| Use ARMesh vertex-based refinement tokenization instead of BPT triangle-based tokenization | Unconditional mesh generation · ShapeNet bench, 50% autoregressive steps | COV: 34.11 -> 54.84 | Helped |
| Use ARMesh vertex-based refinement tokenization instead of BPT triangle-based tokenization | Unconditional mesh generation · ShapeNet bench, 100% autoregressive steps | COV: 56.35 -> 56.29 | No clear effect |
| Use ARMesh vertex-based refinement tokenization instead of BPT triangle-based tokenization | Unconditional mesh generation · ShapeNet bench, 10% autoregressive steps | MMD: 3.02 -> 1.67 | Helped |
| Use ARMesh vertex-based refinement tokenization instead of BPT triangle-based tokenization | Unconditional mesh generation · ShapeNet bench, 20% autoregressive steps | MMD: 1.97 -> 1.15 | Helped |
| Use ARMesh vertex-based refinement tokenization instead of BPT triangle-based tokenization | Unconditional mesh generation · ShapeNet bench, 50% autoregressive steps | MMD: 1.26 -> 0.87 | Helped |
| Use ARMesh vertex-based refinement tokenization instead of BPT triangle-based tokenization | Unconditional mesh generation · ShapeNet bench, 100% autoregressive steps | MMD: 0.76 -> 0.75 | No clear effect |
| Use ARMesh tokenization instead of EdgeRunner tokenization | Unconditional mesh generation · ShapeNet bench, 10% autoregressive steps | COV: 9.43 -> 41.07 | Helped |
| Use ARMesh tokenization instead of EdgeRunner tokenization | Unconditional mesh generation · ShapeNet bench, 10% autoregressive steps | MMD: 10.32 -> 1.67 | Helped |
| Remove the boundary-edge penalty (ablation) | GSlim mesh simplification · Mesh simplification | preservation of holes and leg length | Hurt |
| Apply a vertex penalty (ablation) | GSlim mesh simplification · Mesh simplification | adaptive triangle distribution to geometric curvature | Hurt |
Surprising findings
- Full-generation metrics are only comparable to baselines rather than clearly superior; the advantage is concentrated in early-stopped partial generation.
Limitations
- Limited generalization compared with diffusion-based continuous-space methods, attributed to limited training data.
- Generation complexity is linear in the number of vertex-split steps.
- Topological flexibility is unconstrained and the PSC pipeline lacks gradients, so it cannot be fine-tuned end to end.
Experimental setup
Datasets: ShapeNet Stanford 3D Scans Thingi10K NeRF AMASS
Benchmarks: ShapeNet bench ShapeNet chair ShapeNet table ShapeNet lamp
Metrics: MMD COV 1-NNA FID KID
Baselines: PolyGen MeshGPT PivotMesh EdgeRunner BPT DMesh
Builds on: QSlim PSC (progressive simplicial complex) MeshGPT VAR PolyGen
Closest papers by meaning
- EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation (ICLR 2025 · similarity 0.91)
- MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers (ICLR 2025 · similarity 0.91)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.90)
- PivotMesh: Generic 3D Mesh Generation via Pivot Vertices Guidance (ICLR 2025 · similarity 0.89)
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.89)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.89)
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.