Multi-Modal Foundation · ADSKAILab
Make-A-Shape · single-view 20M
Generates detailed 3D geometric shapes from single 2D images in two seconds for engineering research.
What Make-A-Shape · single-view 20M is
Make-A-Shape single-view 20M is a conditional 3D generative model developed by researchers at the Autodesk AI Lab (ADSKAILab). It generates 3D geometries directly from a single 2D image, capturing complex topologies and fine structural features. It is intended for scientific research and theoretical exploration into shape synthesis and computer-aided design (CAD) reconstruction.
Generating plausible 3D shapes from a single viewpoint is an ill-posed problem due to self-occlusions and missing depth information. Make-A-Shape addresses this challenge by learning geometric priors from large-scale 3D data, producing complete shapes in approximately two seconds.
How it works
The model is built on a U-ViT architecture that incorporates learnable skip-connections between its convolutional and deconvolutional blocks. 3D geometries are represented as truncated signed distance functions (TSDFs) at a resolution of 256³, which are decomposed using a discrete wavelet transform into a wavelet-tree representation. A subband adaptive training strategy is employed to balance the learning of coarse structures and fine details.
Training was conducted across 48 NVIDIA A10G GPUs over approximately 20 days (around 23,000 GPU hours). The model was trained on an aggregated collection of over 10 million publicly available 3D shapes drawn from 18 distinct sources, including ShapeNet, ModelNet, ABC, Fusion 360, BuildingNet, Thingi10K, Infinigen, Objaverse, and subsets of ObjaverseXL.
Key facts
- Model architecture
- U-ViT with learnable skip-connections
- Input condition
- Single 2D image
- Internal representation
- 256³ TSDF wavelet-tree
- Training dataset size
- Over 10 million 3D shapes across 18 sub-datasets
- Inference latency
- Approximately 2 seconds
- Compute used for training
- 48 × A10G GPUs (~23,000 GPU hours)
What you can use it for
- Single-view CAD reconstruction
- Reconstructing three-dimensional geometric structures from individual 2D product photographs or technical illustrations for design exploration.
- Engineering shape synthesis
- Synthesising rapid 3D geometric conceptual models with intricate topology for downstream computational analysis.
- Geometric generative modelling research
- Benchmarking image-to-3d generation performance using standard quantitative 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
- Synthesises 3D shapes from a single-view image within approximately two seconds.
- Retains intricate geometric surface detail and topology using a 256³ wavelet-tree truncated signed distance representation.
- Trained on a corpus of more than 10 million 3D shapes covering diverse categories from CAD components to organic meshes.
- Demonstrates zero-shot generalisation on unseen evaluation datasets including Google Scanned Objects.
Watch out for
- Outputs must not be used for load-bearing physical objects where failure could cause injury or damage.
- Reconstruction fidelity degrades when the input 2D image is ambiguous, occluded, or of low visual quality.
- Geometric accuracy decreases on object classes and styles that are underrepresented within the 18 training datasets.
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 · single-view 20M commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What input does this model require?
The model requires a single 2D image from which it infers and generates the corresponding 3D shape.
What volumetric representation does the model use?
It operates on a truncated signed distance function (TSDF) at 256³ resolution, decomposed via discrete wavelet transform into a wavelet tree.
How quickly does the model generate shapes?
The model generates 3D geometries in approximately two seconds during inference.
Can the generated shapes be used directly in physical manufacturing?
No. Generated shapes are not guaranteed to have real-world structural soundness and must not be used for load-bearing physical components.
How was the model evaluated?
Performance was measured using Intersection over Union (IoU), Chamfer Distance (CD), and Light Field Distance (LFD) on test splits of the training data and the Google Scanned Objects benchmark.
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. In Proceedings of the 41st International Conference on Machine Learning, PMLR 235:20660-20681. arXiv:2401.11067.
Written from Make-A-Shape · single-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.