Research map / 3D vision / Face and shape
MeshXL: Neural Coordinate Field for Generative 3D Foundation Models
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.
Introduces MeshXL, a family of decoder-only auto-regressive transformers (125M–1.3B, OPT-initialized) that generate polygon meshes directly from a discretized coordinate representation (NeurCF), pre-trained on ~2.5M meshes.
Problem
Direct generation of high-fidelity polygon meshes is hard due to unstructured graph representation; prior methods rely on intermediate representations or VQVAE reconstruction that accumulates errors.
Approach
Represent mesh vertices as explicit discretized coordinates embedded with learnable coordinate embeddings (NeurCF), order faces and vertices with a fixed scheme, and model the flattened sequence with next-token prediction in an LLM (OPT) framework, avoiding a mesh VQVAE.
Key results
- On ShapeNet chair, MeshXL (1.3B) reports COV 51.60, MMD 3.23, 1-NNA 55.80, versus MeshGPT COV 42.00, MMD 4.75, 1-NNA 69.50.
- On Objaverse, scaling from 125M to 1.3B improves COV from 39.76 to 42.86 and FID from 17.32 to 12.49.
- User study (434 responses) rates MeshXL quality 3.96 vs. 3.15 for GET3D and 2.53 for PolyGen (reals 4.08).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase model size from 125M to 1.3B (ablation) | MeshXL · Objaverse | COV: 39.76 -> 42.86 | Helped |
| Increase model size from 125M to 1.3B (ablation) | MeshXL · Objaverse | MMD: 5.21 -> 4.16 | Helped |
| Increase model size from 125M to 1.3B (ablation) | MeshXL · Objaverse | 1-NNA: 67.34 -> 61.56 | Helped |
| Increase model size from 125M to 1.3B (ablation) | MeshXL · Objaverse | JSD: 26.03 -> 20.99 | Helped |
| Use direct next-coordinate mesh generation instead of a mesh VQVAE | MeshXL · ShapeNet | COV: 42.00 -> 50.80 | Helped |
Surprising findings
- MeshXL (1.3B) does not consistently beat smaller variants on ShapeNet; e.g., on the Lamp category the 350M model has better MMD and FID than 1.3B.
Limitations
- Slow autoregressive inference (7,200 tokens for an 800-faced mesh).
- Only meshes with fewer than about 800 faces (after decimation) are modeled.
- Larger models do not uniformly improve metrics across all ShapeNet categories.
Experimental setup
Datasets: ShapeNet V2 3D-FUTURE Objaverse Objaverse-XL
Benchmarks: ShapeNet (chair, table, bench, lamp) Objaverse
Metrics: COV MMD 1-NNA JSD FID KID user study scores
Baselines: PolyGen GET3D MeshGPT PolyDiff
Builds on: OPT MeshGPT PolyGen Q-Former
Closest papers by meaning
- ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.87)
- MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers (CVPR 2024 · similarity 0.87)
- PivotMesh: Generic 3D Mesh Generation via Pivot Vertices Guidance (ICLR 2025 · similarity 0.86)
- MeshFormer : High-Quality Mesh Generation with 3D-Guided Reconstruction Model (NeurIPS 2024 · similarity 0.86)
- Implicit-ARAP: Efficient Handle-Guided Neural Field Deformation via Local Patch Meshing (NeurIPS 2025 · similarity 0.85)
- Rigel3D: Rig-aware Latents for Animation-Ready 3D Asset Generation (NeurIPS 2026 · similarity 0.85)
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.