Research map / Point clouds / Shape and surface completion
SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images
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.
SPAR3D is a two-stage single-image 3D reconstructor that first samples a sparse point cloud with a lightweight point diffusion model and then uses a triplane transformer with differentiable PBR rendering to produce a textured mesh conditioned on the point cloud and image.
Problem
Single-image 3D object reconstruction suffers from a trade-off: regression methods align well with visible surfaces but over-smooth occluded regions, while diffusion-based generative methods model uncertainty but are slow and misaligned with the visible input.
Approach
Offload the uncertain (occluded) geometry modeling to a low-resolution point-cloud diffusion stage, then use a regression-based triplane transformer conditioned on the sampled points and image to produce high-fidelity meshes with PBR materials and environment lighting, which also enables interactive point-cloud editing.
Key results
- On GSO, SPAR3D achieves CD 0.120 and FS@0.1 0.584, outperforming the best baseline InstantMesh (CD 0.135) and SF3D (CD 0.137) and CRM (CD 0.161).
- On OmniObject3D, SPAR3D achieves CD 0.122 and FS@0.1 0.587 versus SF3D CD 0.138 and TripoSR CD 0.144.
- SPAR3D runs in 0.7 seconds per object, much faster than multi-view diffusion methods such as InstantMesh (36.1 s) and Era3D (348 s) on GSO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the point sampling stage (ablation) | SPAR3D · GSO | CD: 0.136 -> 0.120 | Helped |
| add the point sampling stage (ablation) | SPAR3D · GSO | FS@0.1: 0.506 -> 0.584 | Helped |
| add the point sampling stage (ablation) | SPAR3D · GSO | PSNR: 18.5 -> 18.6 | Helped |
| add the point sampling stage (ablation) | SPAR3D · GSO | LPIPS: 0.146 -> 0.139 | Helped |
| add the point sampling stage (ablation) | SPAR3D · OmniObject3D | CD: 0.140 -> 0.122 | Helped |
| add the point sampling stage (ablation) | SPAR3D · OmniObject3D | FS@0.1: 0.509 -> 0.587 | Helped |
| add the point sampling stage (ablation) | SPAR3D · OmniObject3D | PSNR: 17.8 -> 17.9 | Helped |
| add the point sampling stage (ablation) | SPAR3D · OmniObject3D | LPIPS: 0.146 -> 0.140 | Helped |
Surprising findings
- SSIM is slightly worse than the strongest baseline, which the authors attribute to Monte Carlo noise from the shader.
- Removing the point sampling stage (regressive variant) causes a significant drop, e.g., FS@0.1 from 0.584 to 0.506 on GSO.
Limitations
- Visible-surface fidelity relies on the meshing stage, and back-surface quality depends on the sampled point cloud quality.
- Released code and weights are only promised upon publication ('We will release our model upon publication').
Experimental setup
Datasets: GSO OmniObject3D
Benchmarks: GSO OmniObject3D
Metrics: Chamfer Distance F-score PSNR SSIM LPIPS inference time
Baselines: Shap-E LN3Diff LGM CRM 3DTopia-XL Era3D TripoSR InstantMesh SF3D
Builds on: DDPM Point-E PointInfinity SF3D RENI++
Closest papers by meaning
- Sparc3D: Sparse Representation and Construction for High-Resolution 3D Shapes Modeling (NeurIPS 2025 · similarity 0.90)
- SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction (CVPR 2025 · similarity 0.90)
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.90)
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · similarity 0.90)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.89)
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · similarity 0.88)
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.