Multi-Modal Foundation · ADSKAILab
Make-A-Shape · multi-view 20M
Generates detailed 3D geometric shapes from four multi-view images for engineering and shape modelling research.
What Make-A-Shape · multi-view 20M is
Make-A-Shape multi-view 20M is a 3D generative model designed to reconstruct complete three-dimensional geometry from multiple 2D viewpoints. Developed by researchers at Autodesk AI Lab, the model targets computer-aided design (CAD) reconstruction, engineering, and materials science applications by synthesising complex topologies and intricate geometric structures.
The model addresses the challenge of recovering high-fidelity 3D geometry from limited visual perspectives. By using a specialised representation that encodes coarse and fine geometric features, it enables fast and structurally plausible shape synthesis across diverse object domains.
Intended strictly for scientific exploration and academic development, the framework provides researchers with a tool to explore automated 3D shape generation, CAD asset synthesis, and visual-to-geometric reconstruction pipelines.
How it works
The model employs a U-ViT (Vision Transformer with U-Net skip connections) architecture incorporating learnable skip connections between its convolution and deconvolution blocks. 3D geometries are represented as truncated signed distance functions (TSDFs) at a resolution of 256³, which are then decomposed using a discrete wavelet transform into a wavelet-tree representation. A subband adaptive training strategy is utilised to capture both coarse structural shapes and high-frequency surface details.
Training was conducted on an aggregated corpus of more than 10 million publicly available 3D shapes from 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. The multi-view model conditions on four view-specific images and synthesises output shapes within approximately two seconds.
Key facts
- Architecture
- U-ViT with learnable skip connections and wavelet-tree representation
- Input modality
- Four view-specific 2D images
- Output modality
- 3D geometric shape (TSDF at 256³ resolution)
- Training data
- Over 10 million 3D shapes from 18 public datasets
- Training compute
- 48 × NVIDIA A10G GPUs (approximately 23,000 GPU hours)
- Inference speed
- Within 2 seconds for most conditions
What you can use it for
- Multi-view CAD shape reconstruction
- Reconstruct 3D geometric CAD representations from four calibrated 2D perspective renderings.
- 3D shape synthesis for engineering design
- Generate diverse structural designs and explore theoretical 3D morphologies from multi-angle image inputs.
- Geometric benchmark evaluation
- Assess image-to-3D generation pipelines across standard geometric metrics such as Intersection over Union and Chamfer Distance.
Run it here
We list this model but cannot run it yet: no library and no architectures — nothing can load it.
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 geometry with complex topologies within two seconds under most conditions.
- Trained on an extensive repository of over 10 million diverse 3D shapes spanning 18 distinct public datasets.
- Demonstrates strong generalisation on unseen shapes, achieving an IoU of 0.7460 and Chamfer Distance of 0.00337 on the Google Scanned Objects benchmark.
- Uses a wavelet-tree representation to balance structural integrity with fine geometric details.
Watch out for
- Output quality relies heavily on the quality, clarity, and unambiguity of the input images.
- Performance may degrade on object categories, styles, or classes underrepresented in the 10-million shape training data.
- Cannot be used to design load-bearing components or physical structures requiring verified real-world structural soundness.
Inputs and outputs
You provide
Image
PNG, JPEG or TIFF. Sent to the model as-is.
You get back
A generated mesh, offered for download — GLB or OBJ.
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 · multi-view 20M commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What inputs does the multi-view model require?
The model takes four view-specific 2D images of an object as its input condition to synthesise the corresponding 3D shape.
How are the output 3D shapes internally represented?
Outputs are derived from a truncated signed distance function (TSDF) at a 256³ resolution, decomposed into a wavelet-tree representation via discrete wavelet transforms.
Can the generated models be directly fabricated for load-bearing applications?
No. The model is intended strictly for theoretical and academic exploration; generated geometries should not be relied upon for physical structural soundness or load-bearing applications.
How quickly can the model reconstruct a 3D geometry?
The model generates 3D shapes within approximately two seconds for most conditions.
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. https://proceedings.mlr.press/v235/hui24a.html
Written from Make-A-Shape · multi-view 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.