Research map / 3D vision / Face and shape
Dora: Sampling and Benchmarking for 3D Shape Variational Auto-Encoders
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 Dora-VAE, a 3D shape VAE that adds sharp edge sampling and dual cross-attention to a Vecset transformer, plus the Dora-Bench benchmark with a Sharp Normal Error metric.
Problem
Uniform point sampling in Vecset-based 3D shape VAEs loses fine geometric detail, limiting reconstruction and downstream diffusion-based 3D generation, while volumetric VAEs need very large latent codes.
Approach
Sample extra points along salient mesh edges (dihedral angle above a threshold) in addition to uniform points, encode the two point sets with separate cross-attentions, and sum them; evaluate on complexity-stratified test sets with a sharp-region normal error metric.
Key results
- With 1280 latent codes, Dora-VAE reduces Chamfer Distance to 2.097 vs 4.015 for fine-tuned XCube-VAE on Dora-Bench (about 47.8% improvement).
- Dora-VAE with 256 latent codes already surpasses fine-tuned XCube-VAE in Chamfer Distance (3.356 vs 4.015).
- On the most complex L4 shapes, Dora-VAE achieves SNE 1.579 vs 1.639 for XCube-VAE†, while using over 8x fewer latent codes than XCube.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sharp edge sampling and dual cross-attention (ablation) | Dora-VAE · Dora-bench L4 | F-score (0.005) × 100: 68.669 -> 87.473 | Helped |
| add sharp edge sampling and dual cross-attention (ablation) | Dora-VAE · Dora-bench L4 | CD × 10000: 24.527 -> 5.265 | Helped |
| add sharp edge sampling and dual cross-attention (ablation) | Dora-VAE · Dora-bench L4 | SNE × 100: 3.035 -> 1.579 | Helped |
| add dual cross-attention (ablation) | Dora-VAE with sharp edge sampling · Dora-bench L4 | F-score (0.005) × 100: 79.240 -> 87.473 | Helped |
| add dual cross-attention (ablation) | Dora-VAE with sharp edge sampling · Dora-bench L4 | CD × 10000: 6.196 -> 5.265 | Helped |
| add dual cross-attention (ablation) | Dora-VAE with sharp edge sampling · Dora-bench L4 | SNE × 100: 2.207 -> 1.579 | Helped |
| use Dora-VAE instead of Xcube† | 3D VAE reconstruction · Dora-bench L4 | F-score (0.005) × 100: 87.365 -> 87.473 | Helped |
| use Dora-VAE instead of Xcube† | 3D VAE reconstruction · Dora-bench L4 | CD × 10000: 7.627 -> 5.265 | Helped |
| use Dora-VAE instead of Xcube† | 3D VAE reconstruction · Dora-bench L4 | SNE × 100: 1.639 -> 1.579 | Helped |
| use Dora-VAE with 256 latent codes instead of Xcube† | 3D VAE reconstruction · Dora-bench L1 | CD × 10000: 4.015 -> 3.356 | Helped |
Surprising findings
- A 256-code Dora-VAE variant without SES and DCA performs far worse than with them, but the 256-code full model still beats the much larger-latent fine-tuned XCube on CD.
Limitations
- Relies on a fixed dihedral-angle threshold to define salient edges.
- Downstream evaluation is limited to a qualitative single-image-to-3D comparison with no quantitative generation metrics.
- Dora-Bench is built on existing public datasets and SNE depends on rendering and Canny edge parameters.
Experimental setup
Datasets: Objaverse GSO ABO Meta
Benchmarks: Dora-Bench
Metrics: F-score (0.01) F-score (0.005) Chamfer Distance Sharp Normal Error latent code length
Baselines: XCube-VAE Craftsman-VAE 3DShape2VecSet
Builds on: 3DShape2VecSet XCube CLAY Poisson disk sampling
Closest papers by meaning
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.87)
- Make-A-Shape: a Ten-Million-scale 3D Shape Model (ICML 2024 · similarity 0.86)
- LaGeM: A Large Geometry Model for 3D Representation Learning and Diffusion (ICLR 2025 · similarity 0.86)
- Sparc3D: Sparse Representation and Construction for High-Resolution 3D Shapes Modeling (NeurIPS 2025 · similarity 0.86)
- ShapeShifter: 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion (CVPR 2025 · similarity 0.85)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.85)
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.