Research map / 3D vision / Face and shape
Sparc3D: Sparse Representation and Construction for High-Resolution 3D Shapes Modeling
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 Sparcubes, a sparse deformable marching cubes remeshing method for watertight 1024^3 surfaces, and Sparconv-VAE, a sparse-convolutional modality-consistent 3D VAE that plugs into TRELLIS latent diffusion for higher-resolution 3D generation.
Problem
High-fidelity 3D shape generation is limited by detail loss from non-watertight mesh conversion (UDF-to-SDF resolution halving, largest-component pruning) and by modality mismatch in two-stage VAE plus latent diffusion pipelines.
Approach
Sparcubes builds a sparse cube grid from active voxels with flood-fill coarse signs and gradient-based vertex deformation, and Sparconv-VAE encodes and decodes this same sparse SDF/deformation format with sparse convolutions and a self-pruning decoder, avoiding global attention and modality conversion.
Key results
- Sparcubes at 1024^3 watertight remeshing achieves CD 1.00 on ABO and 3.01 on Objaverse, beating Dora-wt-1024 (1.07 and 4.35).
- Sparcubes remeshing takes around 30 seconds at 1024 resolution versus around 90 seconds for prior watertight conversion.
- Sparconv-VAE converges in under two days on 32 A100s, about four times faster than TRELLIS and vecset VAEs which need roughly seven days.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · ABO | CD (× 10^4): 1.07 -> 1.00 | Helped |
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · ABO | ANC (× 10^2): 76.94 -> 77.75 | Helped |
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · ABO | F1 (× 10^2): 84.56 -> 85.39 | Helped |
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · Objaverse | CD (× 10^4): 4.25 -> 3.01 | Helped |
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · Objaverse | ANC (× 10^2): 75.77 -> 74.98 | Hurt |
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · Objaverse | F1 (× 10^2): 63.84 -> 65.65 | Helped |
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · Wild | CD (× 10^4): 63.7 -> 0.46 | Helped |
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · Wild | ANC (× 10^2): 78.77 -> 88.58 | Helped |
| Replace Dora-wt with Sparcubes remeshing | Watertight remeshing · Wild | F1 (× 10^2): 65.90 -> 97.06 | Helped |
| Add 2D rendering supervision to Sparconv-VAE (ablation) | Sparconv-VAE · VAE reconstruction test set | reconstruction quality | No clear effect |
Surprising findings
- Adding 2D rendering supervision to the Sparconv-VAE yields negligible improvement.
- Sparcubes at 512 resolution outperforms prior pipeline at 1024 resolution on the reported remeshing metrics.
Limitations
- Sparcubes does not retain original texture information.
- Fully closed meshes with internal structures lose hidden interior elements during remeshing.
- Evaluation relies on a self-assembled Wild dataset and test subsets curated from ABO and Objaverse.
Experimental setup
Datasets: Objaverse Objaverse-XL ABO
Benchmarks: ABO Objaverse Wild
Metrics: Chamfer Distance Absolute Normal Consistency F1 score
Baselines: TRELLIS Craftsman Dora XCubes Dora-wt
Builds on: TRELLIS FlexiCubes Dora Dual Marching Cubes 3DShape2VecSet
Closest papers by meaning
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.90)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.90)
- MeshFormer : High-Quality Mesh Generation with 3D-Guided Reconstruction Model (NeurIPS 2024 · similarity 0.89)
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.88)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- 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.