Multi-Modal Foundation · ADSKAILab
Make-A-Shape · voxel 16³
Generates detailed 3D geometries from coarse 16³ voxel inputs for engineering and materials prototyping.
What Make-A-Shape · voxel 16³ is
Make-A-Shape · voxel 16³ is a conditional 3D generative model developed by researchers at ADSKAILab. It addresses the challenge of producing detailed, structurally plausible 3D shapes from coarse geometric inputs by converting a 16³ voxel grid into a high-quality 3D representation.
The model provides researchers in engineering and materials science with a method for rapid geometric prototyping. By mapping minimal volumetric information to complex topologies and surface details, it supports theoretical exploration of large-scale 3D shape generation.
How it works
The model is built on a U-ViT (vision transformer) architecture featuring learnable skip-connections between convolution and deconvolution blocks. During preprocessing, training shapes are converted into truncated signed distance functions (TSDFs) at a resolution of 256³ and decomposed using a discrete wavelet transform into a wavelet-tree representation. The model applies a subband adaptive training strategy to capture both coarse and fine shape structures across the wavelet hierarchy.
Training was conducted across 48 A10G GPUs over approximately 20 days (around 23,000 GPU hours) using a compiled dataset of over 10 million publicly available 3D shapes. The training corpus aggregates 18 datasets, including ModelNet, ShapeNet, SMPL, Thingi10K, SMAL, COMA, House3D, ABC, Fusion 360, 3D-FUTURE, BuildingNet, DeformingThings4D, FG3D, Toys4K, ABO, Infinigen, Objaverse, and subsets of ObjaverseXL.
Key facts
- Architecture
- U-ViT with learnable skip-connections and wavelet-tree representation
- Training dataset
- Over 10 million shapes from 18 datasets (including ShapeNet, ABC, Fusion 360, and Objaverse)
- Input format
- 16³ voxel grids
- Output format
- 3D shapes (based on 256³ TSDF wavelet representation)
- Evaluation metrics
- IoU: 0.687, LFD: 2266.41
What you can use it for
- Coarse-to-Fine 3D Prototyping
- Synthesising detailed 3D structures from rough 16³ voxel sketches to rapidly test candidate geometries.
- Volumetric Shape Completion
- Reconstructing intricate surface topologies and details from low-resolution volumetric sensor data or spatial occupancy grids.
- Generative Framework Exploration
- Conducting academic research and benchmark comparisons on large-scale conditional 3D shape models.
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
- Generates detailed 3D geometries within approximately two seconds for most conditions.
- Recovers intricate geometric features and complex topologies from coarse 16³ voxel inputs.
- Trained on an extensive repository of over 10 million diverse 3D models across 18 public datasets.
Watch out for
- Performance may degrade on object categories and styles that are underrepresented in the training corpora.
- Output quality is sensitive to input clarity and can yield implausible geometries if inputs are ambiguous.
- Not intended or certified for verifying real-world structural soundness or fabricating load-bearing objects.
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 16³ commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What input does this specific variant require?
This model takes a coarse 16³ voxel grid as input and reconstructs a high-resolution 3D geometric shape.
How quickly can the model generate a 3D geometry?
The model generates shapes in under two seconds for most conditions.
Can the generated geometries be used directly for structural or load-bearing parts?
No. The model is intended for scientific research and theoretical exploration; generated shapes must not be relied upon for structural soundness or load-bearing physical applications.
How was the model evaluated?
It was evaluated using Intersection over Union (IoU) and Light Field Distance (LFD) on a validation set comprising 2% of the shapes from each training sub-dataset, achieving an IoU of 0.687 and an LFD of 2266.41.
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 16³'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.