Research map / 3D vision / 3D generation
MeshArt: Generating Articulated Meshes with Structure-Guided Transformers
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.
Generates articulated 3D meshes part-by-part with two quantized-triangle transformers: first a structure of part boxes with articulation modes, then per-part mesh triangles conditioned on structure and junction faces.
Problem
Generating articulated 3D objects as clean, compact triangle meshes with functional part articulations, which prior methods handle poorly due to retrieval inconsistencies and over-smoothed implicit decoding, and which are limited by scarce articulation annotations.
Approach
Represent both articulation structure (part bounding boxes as triangle meshes with joint attributes) and part meshes as sequences of RVQ-quantized triangle embeddings, generated autoregressively by GPT-style transformers, with part-mesh generation conditioned on the object structure and junction faces from neighboring parts.
Key results
- Average 57.1% improvement in structure coverage over baselines (storage, table, chair structure generation).
- Mesh generation FID improves by 209 points on average over NAP (e.g., table FID 14.3 vs 252.6).
- Storage furniture FID 8.1 for MeshArt vs 170.6 for NAP; ablation removing junction conditioning raises FID from 8.6 to 16.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use triangle-based structure parameterization instead of min/max bounds (ablation) | MeshArt structure generation · Storage Furniture articulated structure | AID: 36.3 -> 39.1 | Helped |
| use triangle-based structure parameterization instead of bounding-box corners (ablation) | MeshArt structure generation · Storage Furniture articulated structure | AID: 35.0 -> 39.1 | Helped |
| add junction-face conditioning (ablation) | MeshArt geometry generation · Storage Furniture articulated mesh | FID: 16.7 -> 8.6 | Helped |
| add shape-structure guidance (ablation) | MeshArt geometry generation · Storage Furniture articulated mesh | FID: 9.6 -> 8.6 | Helped |
| add flexible positional embeddings (ablation) | MeshArt geometry generation · Storage Furniture articulated mesh | FID: 8.8 -> 8.6 | Helped |
| use MeshArt rather than NAP | articulated mesh generation · Chair | FID: 267.7 -> 40.8 | Helped |
| use MeshArt rather than NAP | articulated mesh generation · Table | FID: 252.6 -> 14.3 | Helped |
| use MeshArt rather than NAP | articulated mesh generation · Storage Furniture | FID: 170.6 -> 8.1 | Helped |
Limitations
- Codebook selection is non-differentiable, so physical plausibility and articulation motion are not enforced.
- Articulated PartNet annotations remain limited in size for large-scale training.
- Parts are limited to fewer than 700 faces after decimation to fit the context window.
Experimental setup
Datasets: PartNet PartNet-Mobility
Benchmarks: Articulated PartNet
Metrics: AID ID COV MMD 1-NNA FID KID Chamfer distance novelty
Baselines: NAP CAGE
Builds on: MeshGPT NAP CAGE PolyGen PartNet
Closest papers by meaning
- MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers (ICLR 2025 · similarity 0.91)
- CAGE: Controllable Articulation GEneration (CVPR 2024 · similarity 0.90)
- SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects (ICLR 2025 · similarity 0.90)
- ArtFormer: Controllable Generation of Diverse 3D Articulated Objects (CVPR 2025 · similarity 0.90)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.89)
- Efficient Part-level 3D Object Generation via Dual Volume Packing (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.