Multi-Modal Foundation · ADSKAILab
Make-A-Shape · point-cloud 20M
Generates detailed 3D geometric shapes from point cloud inputs for engineering and computational research.
What Make-A-Shape · point-cloud 20M is
Make-A-Shape point-cloud 20M is a conditional 3D generative model developed by researchers at Autodesk AI Lab (ADSKAILab). It generates detailed, watertight 3D geometries directly from sparse point-cloud inputs, reconstructing intricate structural features and complex topologies.
The model forms part of the broader Make-A-Shape framework, designed to tackle the challenge of high-resolution, structurally plausible 3D shape generation across diverse object categories. By operating on a continuous wavelet representation, it avoids standard computational bottlenecks associated with high-resolution voxel grids while preserving both global form and local surface detail.
Designed primarily for academic and scientific exploration in engineering, graphics, and materials research, the system enables researchers to reconstruct surface geometry and explore shape distributions rapidly without requiring labour-intensive manual modelling.
How it works
The model is built on a U-ViT (Vision Transformer with U-Net skip-connections) architecture, using learnable skip-connections between its convolution and deconvolution blocks. 3D training shapes are converted into truncated signed distance functions (TSDF) at a resolution of 256³ and subsequently decomposed using a discrete wavelet transform into a compact wavelet-tree representation. A subband adaptive training strategy is employed to learn coarse global layouts alongside fine local details effectively.
The training dataset comprises more than 10 million 3D shapes compiled across 18 public repositories, including ModelNet, ShapeNet, SMPL, Thingi10K, SMAL, COMA, House3D, ABC, Fusion 360, 3D-FUTURE, BuildingNet, DeformingThings4D, FG3D, Toys4K, ABO, Infinigen, Objaverse, and subsets of ObjaverseXL. The model was trained on 48 NVIDIA A10G GPUs over approximately 20 days (around 23,000 GPU hours).
Key facts
- Architecture
- U-ViT with learnable skip-connections
- Internal representation
- Wavelet-tree decomposition of 256³ truncated signed distance functions (TSDF)
- Training data
- Over 10 million 3D shapes from 18 datasets
- Inference time
- Within two seconds for most inputs
- Validation IoU (2,500 points)
- 0.7595
- Validation LFD (2,500 points)
- 1857.84
What you can use it for
- Point-Cloud Surface Reconstruction
- Completing full 3D boundary representations and surfaces from unorganised or sparse point-cloud measurements (e.g. 2,500 points).
- CAD and Engineering Geometry Exploration
- Synthesising plausible mechanical and structural shapes from preliminary coordinate samples for design space investigation.
- Shape Distribution Benchmarking
- Evaluating generative reconstruction performance against standards using Intersection over Union (IoU) and Light Field Distance (LFD) metrics.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Rapid generation capability, producing 3D shapes in under two seconds under standard operating conditions.
- Captures complex topologies and sharp geometric details using a multi-scale wavelet-tree formulation.
- Broad geometric prior derived from training on over 10 million varied 3D models spanning synthetic, organic, and architectural domains.
Watch out for
- Output quality is sensitive to the density, clarity, and noise level of the input point cloud.
- Ambiguous or degraded point clouds can lead to the generation of implausible shape artefacts.
- Performance deteriorates for niche object categories or unconventional geometric styles underrepresented in the 18 training datasets.
- Generated models cannot be assumed to possess real-world physical structural soundness or load-bearing integrity.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · other
The other licence carries terms worth reading before you depend on it.
Questions
Can I use Make-A-Shape · point-cloud 20M commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What inputs does this specific variant accept?
It accepts 3D point clouds as conditional inputs, demonstrated and benchmarked on inputs such as 2,500 coordinate points.
What mathematical representation does the model output?
It models shapes via a discrete wavelet transform of a 256³ truncated signed distance function (TSDF) organised into a wavelet-tree.
How quickly can the model generate a 3D shape?
The model generates shapes within two seconds for most operating conditions.
Can the generated geometries be used for physical load-bearing applications?
No. The model is intended strictly for scientific research, and theoretical plausibility in the output does not ensure physical strength or structural reliability.
How to cite it
Hui, K.-H., Sanghi, A., Rampini, A., Rahimi Malekshan, K., Liu, Z., Shayani, H., & Fu, C.-W. (2024). Make-A-Shape: a Ten-Million-scale 3D Shape Model. Proceedings of the 41st International Conference on Machine Learning, PMLR 235:20660-20681.
Written from Make-A-Shape · point-cloud 20M's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.