Research map / Point clouds / Shape and surface completion
TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing
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 TreeMeshGPT, an autoregressive Transformer that generates artistic meshes conditioned on point clouds by traversing triangle adjacency with a DFS stack, using two tokens per face and reaching up to 5,500 faces.
Problem
Autoregressive artistic mesh generation is limited in face count and suffers from flipped normals and mesh artifacts, because prior methods use long token sequences and standard next-token prediction.
Approach
Replaces next-token prediction with Autoregressive Tree Sequencing: the next input edge is popped from a dynamic DFS stack over the half-edge triangle adjacency, so each face adds only two sequence steps and generation extends locally from the last face.
Key results
- On the Objaverse validation set, TreeMeshGPT reaches CD 0.0070 and NC 0.798 versus MeshAnything's 0.0115 / 0.223 and MeshAnythingV2's 0.0102 / 0.167.
- In a controlled tokenizer comparison, the proposed tokenization achieves CD 0.0100 with 2Nf+2Nc sequence length versus 9Nf for naive tokenization (CD 0.0376).
- With 7-bit discretization the model generates meshes with up to 5,500 faces under 2,048 point-cloud tokens; on GSO it achieves CD 0.0077 vs 0.0105 for MeshAnything.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TreeMeshGPT instead of MeshAnythingV2 | Artistic mesh generation conditioned on point clouds · Objaverse evaluation set | CD: 0.0102 -> 0.0070 | Helped |
| Use TreeMeshGPT instead of MeshAnythingV2 | Artistic mesh generation conditioned on point clouds · Objaverse evaluation set | NC: 0.167 -> 0.798 | Helped |
| Use TreeMeshGPT instead of MeshAnythingV2 | Artistic mesh generation conditioned on point clouds · GSO | CD: 0.0116 -> 0.0077 | Helped |
| Use TreeMeshGPT instead of MeshAnythingV2 | Artistic mesh generation conditioned on point clouds · GSO | NC: 0.3269 -> 0.842 | Helped |
| Replace the compared tokenizers with TreeMeshGPT's tokenizer | Controlled mesh-generation experiment · Controlled tokenizer experiment | CD: 0.0327 -> 0.0100 | Helped |
| Replace the compared tokenizers with TreeMeshGPT's tokenizer | Controlled mesh-generation experiment · Controlled tokenizer experiment | NC: -0.069 -> 0.734 | Helped |
| Replace the compared tokenizers with TreeMeshGPT's tokenizer | Controlled mesh-generation experiment · Controlled tokenizer experiment | |NC|: 0.768 -> 0.874 | Helped |
| Use hierarchical coordinate-prediction heads instead of simultaneous x-y-z prediction (ablation) | TreeMeshGPT · Objaverse evaluation set | CD: 0.0114 -> 0.0070 | Helped |
| Use hierarchical coordinate-prediction heads instead of simultaneous x-y-z prediction (ablation) | TreeMeshGPT · Objaverse evaluation set | NC: 0.724 -> 0.798 | Helped |
| Use hierarchical coordinate-prediction heads instead of simultaneous x-y-z prediction (ablation) | TreeMeshGPT · Objaverse evaluation set | |NC|: 0.847 -> 0.880 | Helped |
| Use depth-first traversal instead of breadth-first traversal (ablation) | TreeMeshGPT autoregressive tree sequencing · Training | perplexity | Helped |
Surprising findings
- Predicting coordinates with a naive simultaneous head yields noisier, more incomplete meshes, so the sequential hierarchical head matters for sampling.
- DFS traversal gives better training perplexity than BFS traversal, which the authors attribute to stronger local dependencies.
Limitations
- Success rate decreases as sequence length increases, a failure mode shared with prior methods.
- Enforcing optimal mesh topology remains challenging.
- Quantitative comparisons against baselines use different training datasets and settings, so comparisons are not fully controlled.
Experimental setup
Datasets: Objaverse GSO
Benchmarks: Objaverse validation set (200 samples) GSO
Metrics: Chamfer Distance Normal Consistency absolute Normal Consistency
Baselines: MeshAnything MeshAnythingV2 EdgeRunner PivotMesh VQ-VAE Naive tokenization
Builds on: MeshGPT MeshAnything MeshAnythingV2 EdgeRunner
Closest papers by meaning
- 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 0.88)
- MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers (ICLR 2025 · similarity 0.88)
- Scaling Mesh Generation via Compressive Tokenization (CVPR 2025 · similarity 0.88)
- MeshArt: Generating Articulated Meshes with Structure-Guided Transformers (CVPR 2025 · similarity 0.87)
- MeshGPT: Generating Triangle Meshes with Decoder-Only Transformers (CVPR 2024 · 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.