Multi-Modal Foundation · ADSKAILab
Make-A-Shape · voxel 32³
Generates detailed 3D shapes from 32³ voxel input representations for scientific and engineering research.
What Make-A-Shape · voxel 32³ is
Make-A-Shape (voxel 32³) is a conditional 3D generative model developed by researchers at ADSKAILab. It addresses the challenge of creating detailed, geometrically complex 3D structures from coarse volumetric representations. Given a low-resolution 32³ voxel grid as input, the model produces complete 3D geometries with intricate topological structures.
The model forms part of the broader Make-A-Shape framework, designed to scale 3D generative modelling across diverse shape categories. For researchers in engineering and materials science, it serves as an exploratory tool to convert rough spatial block-outs or low-resolution voxel geometries into high-fidelity candidate shapes within seconds.
How it works
The underlying system uses a U-ViT (Vision Transformer with U-Net skip connections) architecture featuring learnable skip-connections between its convolution and deconvolution blocks. Training relies on an adaptive subband strategy paired with a wavelet-tree representation, allowing the network to capture both coarse structural foundations and fine geometric detail.
During data preparation, training shapes were converted into truncated signed distance functions (TSDF) at a resolution of 256³ and subsequently decomposed using a discrete wavelet transform into wavelet-tree representations. The framework was trained across roughly 23,000 GPU hours on 48 NVIDIA A10G GPUs using an aggregated dataset of over 10 million publicly available 3D shapes spanning 18 sub-datasets (including ShapeNet, ModelNet, ABC, Fusion 360, and Objaverse).
Key facts
- Architecture
- U-ViT with learnable skip-connections and wavelet-tree representation
- Input format
- 32³ voxel grid
- Internal shape representation
- 256³ Truncated Signed Distance Function (TSDF) decomposed via discrete wavelet transform
- Training data
- Over 10 million shapes from 18 datasets (including ModelNet, ShapeNet, SMPL, Thingi10K, ABC, Fusion 360, and Objaverse)
- Compute infrastructure
- 48 × NVIDIA A10G GPUs (approx. 20 days / 23,000 GPU hours)
- Inference speed
- Within two seconds for most conditions
- Evaluation metrics (Our Val)
- Light Field Distance (LFD): 1580.98; Intersection over Union (IoU): 0.7942
What you can use it for
- Low-Resolution Shape Exploration
- Transforming coarse 32³ voxel approximations into high-detail continuous 3D geometries during early-stage geometric ideation.
- Voxel-to-3D Geometry Reconstruction
- Synthesising plausible, topologically complex surface boundaries and volumes from low-fidelity spatial occupancy grids.
- Generative Framework Benchmarking
- Evaluating conditional 3D generative performance using standardised geometric metrics like Intersection over Union (IoU) and Light Field Distance (LFD).
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
- Synthesises complete 3D geometries from 32³ voxel inputs within two seconds for most conditions.
- Trained on a diverse collection of over 10 million 3D shapes spanning engineering, architectural, organic, and everyday objects.
- Employs a wavelet-tree representation and subband adaptive training to capture both coarse topological structure and fine boundary details.
- Achieves an Intersection over Union (IoU) of 0.7942 and a Light Field Distance (LFD) of 1580.98 on validation shapes.
Watch out for
- Output quality is directly dependent on the clarity and quality of the input voxel grid, with ambiguous inputs risking implausible structures.
- Performance may degrade on shape categories or object styles that are underrepresented in the 10-million-shape training corpus.
- Outputs must not be relied upon for structural soundness or physical load-bearing applications.
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 · voxel 32³ commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What input format does this model expect?
It accepts 3D voxel representations at a 32³ grid resolution.
How quickly can the model generate a 3D geometry?
The model generates shapes within two seconds under most conditions.
Can the generated geometries be used for physical engineering parts?
No. The model is intended for scientific research and theoretical exploration; generated shapes should not be relied upon for structural integrity or used to build load-bearing physical objects.
What training corpus was used for the model?
It was trained on over 10 million shapes gathered across 18 public datasets, including ModelNet, ShapeNet, ABC, Fusion 360, Thingi10K, BuildingNet, and Objaverse.
How to cite it
Hui, Ka-Hei and Sanghi, Aditya and Rampini, Arianna and Rahimi Malekshan, Kamal and Liu, Zhengzhe and Shayani, Hooman and Fu, Chi-Wing. Make-A-Shape: a Ten-Million-scale 3D Shape Model. Proceedings of the 41st International Conference on Machine Learning, PMLR 235:20660-20681, 2024. arXiv:2401.11067.
Written from Make-A-Shape · voxel 32³'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.