Research map / 3D vision / Face and shape
Learning Versatile 3D Shape Generation with Improved Auto-regressive Models
Simian Luo, Xuelin Qian, Yanwei Fu, Yinda Zhang, Ying Tai, Zhenyu Zhang, Chengjie Wang, Xiangyang Xue
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 coupling network to project three planes into a latent vector before quantization (ablation) | ImAM · ShapeNet plane | 1-NNA: 73.67 -> 59.95 | Improved |
| add coupling network to project three planes into a latent vector before quantization (ablation) | ImAM · ShapeNet plane | ECD: 743 -> 236 | Improved |
| quantize latent vector instead of volumetric grid (ablation) | ImAM · ShapeNet reconstruction | IoU: 88.87 -> 88.01 | No clear effect |
| quantize latent vector instead of volumetric grid (transformer trainability) (ablation) | ImAM · ShapeNet generation stage 2 | trainability | Improved |
| increase feature resolution from 16 to 32 (ablation) | ImAM stage 1 · ShapeNet reconstruction | IoU: 79.17 -> 88.01 | Improved |
| increase codebook entries from 1024 to 4096 (ablation) | ImAM stage 1 · ShapeNet reconstruction | IoU: 86.57 -> 88.01 | Improved |
| use ImAM instead of GBIF for unconditional generation | ImAM · ShapeNet unconditional (5 categories, avg) | ECD: 858 -> 240 | Improved |
| use ImAM vs PVD on Car category | ImAM · ShapeNet unconditional (Car) | 1-NNA: 57.37 (PVD) vs 76.58 | Hurt |
| use ImAM instead of GBIF on COV | ImAM · ShapeNet unconditional (5 categories, avg) | COV: 80.57 -> 78.29 | Hurt |
| use ImAM instead of AutoSDF on bottom-half completion MMD | ImAM · Partial point completion, bottom-half (avg) | MMD (x10^3): 1.751 -> 1.837 | Hurt |
| use ImAM instead of CLIP-Forge for prompt-based text-to-shape FPD | ImAM · Text-guided generation, prompts (Text2Shape) | FPD: 4.14 -> 5.39 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AutoSDF — AutoSDF: Shape Priors for 3D Completion, Reconstruction and Generation (CVPR 2022) · better on 3 of 4 shared results
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 0.90)
- 3DILG: Irregular Latent Grids for 3D Generative Modeling (NeurIPS 2022 · similarity 0.90)
- AutoSDF: Shape Priors for 3D Completion, Reconstruction and Generation (CVPR 2022 · similarity 0.89)
- EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation (ICLR 2025 · similarity 0.88)
- AutoPartGen: Autoregressive 3D Part Generation and Discovery (NeurIPS 2025 · similarity 0.87)
- LaGeM: A Large Geometry Model for 3D Representation Learning and Diffusion (ICLR 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.
