Research map / Point clouds / Shape and surface completion
EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation
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 EdgeRunner, an auto-regressive auto-encoder with an EdgeBreaker-based mesh tokenizer that generates artistic meshes up to 4,000 faces at 512^3 quantization, plus a latent diffusion model for image-conditioned mesh generation.
Problem
Auto-regressive mesh generation is limited to few faces (under ~1,600), coarse 128^3 vertex quantization, and poor generalization, especially when conditioning on single-view images.
Approach
Adapts the EdgeBreaker triangle-traversal compression into a 1D token scheme (about 4-5 tokens per face, versus 9 naive) with no long-range branching, and uses an auto-regressive auto-encoder to map variable-length meshes to a fixed-length latent code on which a DiT latent diffusion model is trained for image conditioning.
Key results
- Tokenizer achieves a 47.4% compression ratio with 54.7 sub-sequences per mesh versus 199.5 for AMT, and is about 3x faster in tokenization (25.2 vs 8.6 meshes per second).
- User study (1-5 scale) gives EdgeRunner 4.58 overall quality versus 2.43 for MeshAnything and 2.14 for MeshAnythingV2.
- Generates meshes up to 4,000 faces at 512^3 quantization; about 45 s for 1,000 faces and about 3 min for 4,000 faces on one A100 at about 100 tokens/s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the EdgeRunner tokenizer instead of AMT | Mesh tokenization · Mesh tokenizer comparison | Sub-sequence Count: 199.5 -> 54.7 | Helped |
| Use the EdgeRunner tokenizer instead of AMT | Mesh tokenization · Mesh tokenizer comparison | Tokenization Speed: 8.6 -> 25.2 | Helped |
| Use the EdgeRunner tokenizer instead of AMT | Mesh tokenization · Mesh tokenizer comparison | Compression Ratio: 46.2% -> 47.4% | No clear effect |
| Use EdgeRunner instead of MeshAnything | Point-conditioned mesh generation · Point-conditioned mesh generation | Input Consistency: 2.93 -> 4.83 | Helped |
| Use EdgeRunner instead of MeshAnythingV2 | Point-conditioned mesh generation · Point-conditioned mesh generation | Input Consistency: 2.64 -> 4.83 | Helped |
| Use EdgeRunner instead of MeshAnything | Point-conditioned mesh generation · Point-conditioned mesh generation | Triangle Aesthetic: 2.43 -> 4.54 | Helped |
| Use EdgeRunner instead of MeshAnythingV2 | Point-conditioned mesh generation · Point-conditioned mesh generation | Triangle Aesthetic: 2.36 -> 4.54 | Helped |
| Use EdgeRunner instead of MeshAnything | Point-conditioned mesh generation · Point-conditioned mesh generation | Overall Quality: 2.43 -> 4.58 | Helped |
| Use EdgeRunner instead of MeshAnythingV2 | Point-conditioned mesh generation · Point-conditioned mesh generation | Overall Quality: 2.14 -> 4.58 | Helped |
| Increase quantization resolution from 128^3 to 512^3 (ablation) | EdgeRunner · Mesh generation quality | Surface quality | Helped |
| Use direct image conditioning instead of ArAE latent conditioning with latent diffusion (ablation) | EdgeRunner image-conditioned generation · Unseen test cases | Generalization | Hurt |
Surprising findings
- Direct image-conditioned auto-regressive decoding converged more slowly and generalized worse than the latent diffusion route, motivating the ArAE design.
- The 128^3 quantization was easier to train for the classification head but showed noticeable quality degradation.
Limitations
- Compression ratio remains insufficient for complex meshes such as game characters, which exceed the face capacity.
- Success rate drops and generation time grows as sequence length increases.
- Auto-regressive sampling yields diverse but sometimes unsatisfactory topologies, and precise control is difficult.
Experimental setup
Datasets: Objaverse Objaverse-XL G-Objaverse
Metrics: compression ratio sub-sequence count tokenization speed user study rating (input consistency, triangle aesthetic, ov
Baselines: MeshAnything MeshAnythingV2 MeshGPT MeshXL Unique3D AMT
Builds on: EdgeBreaker OPT MeshGPT DDPM DDIM
Closest papers by meaning
- ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction (NeurIPS 2025 · similarity 0.91)
- TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing (CVPR 2025 · similarity 0.88)
- MAR-3D: Progressive Masked Auto-regressor for High-Resolution 3D Generation (CVPR 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 0.87)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.87)
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.