Multi-Modal Foundation · ADSKAILab
WaLa PC 1B
Generates complete 3D shapes from point cloud inputs using wavelet-based latent diffusion in 2 to 4 seconds.
What WaLa PC 1B is
WaLa PC 1B is a 3D generative model developed by researchers at Autodesk AI Lab. It forms part of the Wavelet Latent Diffusion (WaLa) framework and is designed to reconstruct complete, geometrically detailed 3D structures from point cloud coordinates.
Deriving solid 3D surfaces from point clouds is a standard challenge in engineering and computational design. WaLa PC 1B addresses this by employing a compact wavelet-tree representation that captures both global shape and fine surface detail, producing high-fidelity outputs in two to four seconds.
How it works
The model represents 3D shapes as truncated signed distance functions (TSDF) at a resolution of 256³. A TSDF records the distance from any point on a voxel grid to the nearest surface boundary. These volumetric grids are decomposed with a discrete wavelet transform into a compact wavelet-tree representation.
Conditioning is driven by point clouds. Input points (sampled at 2,500 points during training) are processed by a PointNet encoder and aggregated via attention pooling into condition latent vectors. A modified U-ViT (a vision Transformer with U-Net skip connections) comprising roughly 967 million parameters then carries out diffusion in the wavelet latent space. The model was trained on NVIDIA H100 GPUs using more than 10 million shapes collected across 19 public repositories, including ShapeNet, ModelNet, ABC, Fusion 360, and Objaverse.
Key facts
- Architecture
- Modified U-ViT with PointNet conditioning encoder
- Parameter count
- Approximately 967 million
- Conditioning input
- Point cloud (2,500 points)
- Internal representation
- 256³ TSDF decomposed via discrete wavelet transform
- Inference speed
- 2 to 4 seconds
- Training compute
- NVIDIA H100 GPUs
- Training dataset
- Over 10 million shapes from 19 public 3D datasets
What you can use it for
- Point Cloud Surface Reconstruction
- Synthesising complete 3D geometries from sparse point cloud coordinate inputs.
- CAD and Mechanical Shape Generation
- Reconstructing solid CAD-style parts from point sets, drawing on training across mechanical datasets such as ABC and Fusion 360.
- Fast Geometric Prototyping
- Generating complex 3D shape candidates within two to four seconds for computational design workflows.
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
- Produces completed 3D shapes rapidly, with inference taking between two and four seconds.
- Uses compact wavelet latents to retain both overall structure and high-frequency surface detail at 256³ resolution.
- Trained on a diverse corpus of more than 10 million shapes aggregated across 19 datasets.
- Achieves an Intersection over Union of 0.9389 and a Chamfer Distance of 0.0011 on the Google Scanned Objects benchmark.
Watch out for
- Reconstruction fidelity degrades when given sparse, incomplete, or noisy input point clouds.
- Generated shapes may lack real-world structural plausibility and must not be used for load-bearing mechanical applications.
- Reflects representation biases from its training datasets, showing weaker performance on rare or unrepresented object categories.
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 WaLa PC 1B commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What conditioning input does the model expect?
The model expects a point cloud. In training, 2,500 points were randomly sampled from each 3D shape.
How quickly can WaLa PC 1B generate a shape?
Inference takes approximately two to four seconds to produce a 3D shape.
How are 3D volumes encoded by the model?
Shapes are first converted into a truncated signed distance function at 256³ resolution and then decomposed into a wavelet-tree representation via discrete wavelet transforms.
Can generated parts be manufactured directly for load-bearing uses?
No. The model can output structurally implausible geometry, and its outputs cannot be relied upon for structural soundness or physical load-bearing tasks.
How to cite it
@misc{sanghi2024waveletlatentdiffusionwala, title={Wavelet Latent Diffusion (Wala): Billion-Parameter 3D Generative Model with Compact Wavelet Encodings}, author={Aditya Sanghi and Aliasghar Khani and Pradyumna Reddy and Arianna Rampini and Derek Cheung and Kamal Rahimi Malekshan and Kanika Madan and Hooman Shayani}, year={2024}, eprint={2411.08017}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2411.08017}}
Written from WaLa PC 1B'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.