Research map / Point clouds / Shape and surface completion
Scaling Mesh Generation via Compressive Tokenization
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 Blocked and Patchified Tokenization (BPT), which compresses mesh token sequences by ~75% via block-wise indexing and face patch aggregation, and uses it to train a point-cloud/image-conditioned autoregressive mesh generator on meshes up to 8k faces.
Problem
Autoregressive mesh generation is limited by long token sequences from vanilla mesh tokenization, which restricts training to low-poly meshes (about 4k faces or fewer) and hurts detail and robustness.
Approach
Block-wise indexing splits vertex coordinates into blocks with per-block offsets to shrink vocabulary and sequence length, and patchified aggregation groups faces around high-degree center vertices with dual-block indexes that replace separator tokens, so sequences are about 75% shorter and meshes beyond 8k faces become trainable.
Key results
- BPT achieves a compression ratio of 0.74 versus 0.54 for MeshAnythingv2 and 0.53 for EdgeRunner (Table 1).
- On point-cloud conditioned generation, the model reaches HD 0.166 and CD 0.094, against 0.265/0.114 for MeshAnythingv2 and 0.301/0.136 for MeshAnything (Table 2).
- Training on meshes with more faces steadily improves Hausdorff distance, and an 8k-face setting outperforms truncated 4k and 8k context training (Figures 2, 7, 8).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BPT instead of MeshAnythingV2 | Point-cloud conditional generation · Point-cloud conditional generation | Hausdorff distance: 0.265 -> 0.166 | Helped |
| use BPT instead of MeshAnything | Point-cloud conditional generation · Point-cloud conditional generation | Hausdorff distance: 0.301 -> 0.166 | Helped |
| use BPT instead of MeshAnythingV2 | Point-cloud conditional generation · Point-cloud conditional generation | Chamfer distance: 0.114 -> 0.094 | Helped |
| use BPT instead of MeshAnything | Point-cloud conditional generation · Point-cloud conditional generation | Chamfer distance: 0.136 -> 0.094 | Helped |
| use BPT instead of MeshAnythingV2 tokenization | Mesh tokenization · Mesh tokenization | Compression ratio: 0.54 -> 0.74 | Helped |
| use block size 4 and offset size 32 instead of block size 8 and offset size 16 (ablation) | BPT · Point-cloud conditional generation | Hausdorff distance: 0.166 -> 0.209 | Hurt |
| use block size 16 and offset size 8 instead of block size 8 and offset size 16 (ablation) | BPT · Point-cloud conditional generation | Hausdorff distance: 0.166 -> 0.256 | Hurt |
| use block size 4 and offset size 32 instead of block size 8 and offset size 16 (ablation) | BPT · Point-cloud conditional generation | Chamfer distance: 0.094 -> 0.111 | Hurt |
| use block size 16 and offset size 8 instead of block size 8 and offset size 16 (ablation) | BPT · Point-cloud conditional generation | Chamfer distance: 0.094 -> 0.126 | Hurt |
| train with meshes having more faces (up to 8000 rather than 1600) (ablation) | Mesh generation model · Mesh generation | Hausdorff distance | Helped |
| train on truncated mesh sequences and use sliding-window attention (ablation) | Mesh generation model · Mesh generation | generation robustness | Hurt |
Surprising findings
- Truncating long mesh sequences to fit smaller context windows with sliding-window inference reduces robustness and produces more incomplete meshes.
Limitations
- Model scale of 500M parameters is considered insufficient; larger scaling not yet explored.
- Topology of generated meshes is not guaranteed and may need post-processing such as mesh fixing.
- Baseline EdgeRunner had no public code, so only qualitative comparison was possible.
Experimental setup
Datasets: ShapeNetV2 3D-FUTURE Objaverse Objaverse-XL internal licensed 3D data
Benchmarks: point-cloud conditioned mesh generation test set of about 500 dense meshes
Metrics: Hausdorff Distance Chamfer Distance compression ratio AVD@t
Baselines: MeshAnything MeshAnythingv2 EdgeRunner
Builds on: MeshXL Michelangelo MeshAnything MeshAnythingv2 adjacent mesh tokenization
Closest papers by meaning
- FreeMesh: Boosting Mesh Generation with Coordinates Merging (ICML 2025 · similarity 0.89)
- TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing (CVPR 2025 · similarity 0.88)
- ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.88)
- EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation (ICLR 2025 · similarity 0.87)
- MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers (CVPR 2024 · similarity 0.86)
- Mesh-RFT: Enhancing Mesh Generation via Fine-grained Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.86)
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.