Research map / Point clouds / Shape and surface completion
Make-A-Shape: a Ten-Million-scale 3D Shape Model
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.
Make-A-Shape trains a large diffusion-based 3D shape generator on a wavelet-tree representation of TSDFs from a 10M-shape corpus, with unconditional and image, point cloud, and voxel conditioned variants.
Problem
Training large-scale 3D generative models is limited by high compute and storage costs, and by representations that are either lossy or inefficient to diffuse.
Approach
A wavelet-tree representation of high-resolution TSDFs with coefficient filtering keeps most shape information compactly; coefficient packing turns it into a low-resolution grid for diffusion, and subband adaptive loss balances coarse and sparse detail coefficients.
Key results
- Single-view model reaches IoU 0.5004 and LFD 3406.61 on GSO, versus Shap-E IoU 0.3488 and One-2-3-45 IoU 0.4159.
- Multi-view model reaches IoU 0.7460 and LFD 1890.85 on GSO, far above single-view results.
- Wavelet-tree representation reconstructs 256^3 SDFs at mean IoU 0.9956, compared with 0.8642 for Point-E's representation and 0.8576 for Shap-E's.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the wavelet-tree representation instead of the coarse component | Make-A-Shape shape encoding · GSO shape reconstruction | IoU: 0.9531 -> 0.9956 | Helped |
| Use multi-view images rather than a single-view image | Make-A-Shape · GSO | IoU: 0.5004 -> 0.7460 | Helped |
| Use multi-view images rather than a single-view image | Make-A-Shape · Our Val | IoU: 0.4285 -> 0.6707 | Helped |
| Replace trilinear voxel upsampling with generative voxel conditioning | Make-A-Shape · Voxel-to-3D, 16^3 | IoU: 0.2373 -> 0.687 | Helped |
| Replace trilinear voxel upsampling with generative voxel conditioning | Make-A-Shape · Voxel-to-3D, 32^3 | IoU: 0.4719 -> 0.7942 | Helped |
| Increase point-cloud input from 2500 to 25000 points (ablation) | Make-A-Shape point-conditioned model · Point-to-3D | IoU: 0.7595 -> 0.8535 | Helped |
| Add subband adaptive training for detail coefficients (ablation) | Make-A-Shape multi-view model · Multi-view generation | IoU: 0.5474 -> 0.6105 | Helped |
| Add subband adaptive training for detail coefficients (ablation) | Make-A-Shape multi-view model · Multi-view generation | LFD: 3191.49 -> 2611.60 | Helped |
| Replace uniform MSE with separate subband MSE losses (ablation) | Make-A-Shape multi-view model · Multi-view generation | IoU: 0.5474 -> 0.5898 | Helped |
| Replace separate subband MSE losses with subband adaptive training (ablation) | Make-A-Shape multi-view model · Multi-view generation | IoU: 0.5898 -> 0.6105 | Helped |
| Increase inference steps from 10 to 100 (ablation) | Make-A-Shape single-view model · Our Val | LFD: 4312.23 -> 4071.33 | Helped |
| Increase inference steps from 10 to 100 (ablation) | Make-A-Shape unconditional model · Our Val | FID: 371.32 -> 85.25 | Helped |
Surprising findings
- Using only the coarse wavelet component yields plausible results, and naive uniform MSE over all subbands degrades quality compared with the coarse-only target.
- Point cloud conditioning performs best with a lower guidance weight (1.0), unlike text-to-image settings that need larger weights.
Limitations
- Multi-view conditioning requires four views with known camera poses, so single-view use depends on separate novel-view models.
- Detail coefficient filtering uses a fixed top-K selection and empirical thresholds; the paper does not explore more sophisticated compression.
Experimental setup
Datasets: ModelNet ShapeNet SMLP Thingi10K SMAL COMA House3D ABC Fusion 360 3D-FUTURE BuildingNet DeformingThings4D
Benchmarks: Google Scanned Objects Our Val (held-out split of compiled dataset)
Metrics: IoU LFD FID
Baselines: Point-E Shap-E One-2-3-45 Nearest-neighbor voxel upsampling Trilinear voxel upsampling Coarse-component-only wavelet generation
Builds on: neural wavelet representation DDPM U-ViT PointNet Set Transformer
Closest papers by meaning
- Mosaic-SDF for 3D Generative Models (CVPR 2024 · similarity 0.89)
- ShapeShifter: 3D Variations Using Multiscale and Sparse Point-Voxel Diffusion (CVPR 2025 · similarity 0.88)
- GPLD3D: Latent Diffusion of 3D Shape Generative Models by Enforcing Geometric and Physical Priors (CVPR 2024 · similarity 0.87)
- Direct3D-S2: Gigascale 3D Generation Made Easy with Spatial Sparse Attention (NeurIPS 2025 · similarity 0.87)
- Large-Vocabulary 3D Diffusion Model with Transformer (ICLR 2024 · similarity 0.87)
- Structured 3D Latents for Scalable and Versatile 3D Generation (CVPR 2025 · similarity 0.86)
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.