Research map / 3D vision / Face and shape
MeshAnything: Artist-Created Mesh Generation with Autoregressive 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.
Proposes MeshAnything, which reframes mesh extraction as shape-conditioned generation of artist-created meshes, using a VQ-VAE with a noise-resistant shape-conditioned decoder and a point-cloud-conditioned OPT-350M transformer.
Problem
Meshes extracted from 3D representations (reconstruction, generation, scanning) are dense and have poor topology, with excessive faces and poor sharp-edge and flat-surface fidelity, making them unsuitable for 3D industry pipelines.
Approach
Treat mesh extraction as generation: an autoregressive transformer conditioned on point clouds (with normals) produces artist-created meshes aligned to the input shape. Training shapes are corrupted via SDF-to-coarse-mesh resampling to close the train-inference domain gap, and the VQ-VAE decoder is fine-tuned to take the shape condition and Gumbel noise on tokens for robustness.
Key results
- On Objaverse, MeshAnything reaches COV 53.1, MMD 2.72, 1-NNA 55.7, FID 14.5 versus MeshGPT's 41.7, 3.83, 67.3, 25.1.
- MeshAnything uses about 0.318K faces on average versus 146K for Marching Cubes, hundreds of times fewer.
- In a 41-user study (1,230 comparisons), MeshAnything won 44.6% on shape and 83.6% on topology against mesh extraction baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace MeshGPT with shape-conditioned MeshAnything | Mesh generation on Objaverse · Objaverse | COV: 41.7 -> 53.1 | Helped |
| Replace MeshGPT with shape-conditioned MeshAnything | Mesh generation on Objaverse · Objaverse | MMD: 3.83 -> 2.72 | Helped |
| Replace MeshGPT with shape-conditioned MeshAnything | Mesh generation on Objaverse · Objaverse | FID: 25.1 -> 14.5 | Helped |
| Replace FlexiCubes with MeshAnything | Mesh extraction · Mesh extraction evaluation | #F: 121.1 -> 0.318 | Helped |
| Replace FlexiCubes with MeshAnything | Mesh extraction · Mesh extraction evaluation | CD: 1.190 -> 2.256 | Hurt |
| Add the noise-resistant decoder (ablation) | MeshAnything · Mesh extraction evaluation | CD: 2.423 -> 2.256 | Helped |
| Add the noise-resistant decoder (ablation) | MeshAnything · Reconstruction, noise level 0.1 | CD: 0.187 -> 0.028 | Helped |
| Increase decoder input token noise from 0.0 to 1.0 while using the noise-resistant decoder (ablation) | MeshAnything decoder · Reconstruction | CD: 0.007 -> 1.798 | Hurt |
| Increase input point-cloud noise from scale 0.005 to 0.050 (ablation) | MeshAnything · Input point cloud quality evaluation | CD: 2.351 -> 4.910 | Hurt |
| Condition MeshAnything on Rodin-generated shapes | MeshAnything · Rodin generated point cloud | CD: 2.552 | Mixed |
Surprising findings
- MeshAnything sometimes produces meshes with fewer faces than the ground-truth artist meshes while matching shape.
- Reconstruction-style baselines (Marching Cubes, FlexiCubes) still score better on CD and NC than MeshAnything.
Limitations
- Cannot generate meshes exceeding the maximum face count limit, so large scenes and very complex objects are unsupported.
- Generative nature makes it less stable and less robust than reconstruction-based extraction, lowering CD and NC.
- Trained only on filtered meshes under 800 faces.
Experimental setup
Datasets: Objaverse ShapeNet
Benchmarks: Objaverse evaluation split ShapeNet
Metrics: COV MMD 1-NNA FID KID CD ECD NC #F user preference votes
Baselines: PolyGen MeshGPT Marching Cubes FlexiCubes Shape As Points Blender remesh
Builds on: MeshGPT PolyGen VQ-VAE Point-BERT
Closest papers by meaning
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.91)
- ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.91)
- MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers (CVPR 2024 · similarity 0.89)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.89)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.89)
- TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing (CVPR 2025 · similarity 0.88)
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.