Research map / 3D vision / Face and shape
MeshWeaver: Sparse-Voxel-Guided Surface Weaving for Autoregressive Mesh Generation
Jiale Xu, Wang Zhao, Ying Shan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add voxel features as vertex representation (ablation) | MeshWeaver · Toys4K | CD (x10^-1): 0.142 -> 0.116 | Improved |
| add cross-attention-guided token prediction (ablation) | MeshWeaver · Toys4K | CD (x10^-1): 0.146 -> 0.116 | Improved |
| add both voxel features and cross-attention (ablation) | MeshWeaver · Toys4K | CD (x10^-1): 0.158 -> 0.116 | Improved |
| add sparse-voxel scaffold masking at inference (ablation) | MeshWeaver · Toys4K | CD (x10^-1): 0.122 -> 0.116 | Improved |
| add sparse-voxel scaffold masking at inference (ablation) | MeshWeaver · Toys4K | HD: 0.090 -> 0.087 | Improved |
| add voxel features as vertex representation (ablation) | MeshWeaver · Toys4K | NC: 0.694 -> 0.732 | Improved |
| use two-level (16,8) instead of three-level (8,4,4) space partition (ablation) | MeshWeaver · Toys4K | CD (x10^-1): 0.137 -> 0.116 | Improved |
| use (16,8) instead of (8,16) space partition (ablation) | MeshWeaver · Toys4K | CD (x10^-1): 0.120 -> 0.116 | No clear effect |
| allocate 20 instead of 16 layers to level 0 (ablation) | MeshWeaver · Toys4K | CD (x10^-1): 0.116 -> 0.121 | No clear effect |
| allocate 18 instead of 16 layers to level 0 (ablation) | MeshWeaver · Toys4K | CD (x10^-1): 0.116 -> 0.113 | No clear effect |
| add cross-attention KV cache (ablation) | MeshWeaver · Toys4K (200 meshes) | throughput (tokens/s): 26.8 -> 30.7 | Improved |
| use vertex-level tokenization instead of coordinate-level | MeshWeaver · mesh tokenization | compression ratio: 0.22 -> 0.18 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- MeshAnythingV2 — MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers (ICLR 2025) · better on 5 of 5 shared results
- Nautilus — Nautilus: Locality-aware Autoencoder for Scalable Mesh Generation (ICCV 2025) · better on 1 of 1 shared result
- DeepMesh — DeepMesh: Differentiable Iso-Surface Extraction (TPAMI 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- 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 0.90)
- MeshFormer : High-Quality Mesh Generation with 3D-Guided Reconstruction Model (NeurIPS 2024 · similarity 0.89)
- MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers (ICLR 2025 · similarity 0.89)
- PivotMesh: Generic 3D Mesh Generation via Pivot Vertices Guidance (ICLR 2025 · similarity 0.88)
- EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation (ICLR 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.
