Multi-Modal Foundation · ADSKAILab
WaLa VX16 1B
A 1-billion-parameter 3D generative model that reconstructs detailed shapes from 16³ voxel grids in 2–4 seconds.
What WaLa VX16 1B is
WaLa VX16 1B is a 3D generative model designed to generate detailed, geometrically complex 3D shapes conditioned on coarse 16³ voxel grids. Developed by researchers at ADSKAILab (Autodesk AI Lab), the model is a variant of the Wavelet Latent Diffusion (WaLa) framework, scaled to nearly one billion parameters.
Generating detailed 3D structures from coarse spatial inputs addresses a common challenge in computational design, engineering, and materials science, where detailed geometry must often be inferred from low-resolution spatial block-outs or early conceptual volumes. WaLa VX16 1B maps a coarse spatial arrangement into a coherent 3D shape in approximately 2 to 4 seconds.
The model relies on a compact wavelet-based latent space that captures both global structure and fine surface features, enabling rapid high-resolution shape synthesis without requiring excessively dense input representations.
How it works
WaLa VX16 1B uses a modified U-ViT (U-Net Vision Transformer) architecture operating within a compact wavelet-based latent representation. During training, high-resolution 3D shapes are transformed into truncated signed distance functions (TSDF) at a resolution of 256³ and subsequently decomposed using a discrete wavelet transform to form a wavelet-tree representation. Coarse conditioning is derived by downsampling each shape into a 16³ voxel grid, which is processed through a ResNet-based convolutional encoder and downsampled into conditioning latent vectors.
The system was trained on NVIDIA H100 GPUs using a compilation of over 10 million publicly available 3D shapes drawn from 19 datasets, including ModelNet, ShapeNet, SMLP, Thingi10K, SMAL, COMA, House3D, ABC, Fusion 360, 3D-FUTURE, BuildingNet, DeformingThings4D, FG3D, Toys4K, ABO, Infinigen, Objaverse, and subsets of ObjaverseXL.
Key facts
- Model architecture
- Modified U-ViT with ResNet voxel encoder
- Parameter count
- Approximately 906.9 million
- Input format
- 16³ voxel grid
- Target representation
- 256³ Truncated Signed Distance Function (TSDF) wavelet-tree
- Inference time
- 2 to 4 seconds
- Training compute
- NVIDIA H100 GPUs
- Training dataset size
- Over 10 million 3D shapes across 19 sub-datasets
What you can use it for
- Coarse-to-fine shape generation
- Synthesising detailed 3D geometric surfaces from low-resolution 16³ voxel arrangements for rapid geometric prototyping.
- Geometric reconstruction research
- Benchmarking and exploring latent diffusion models and wavelet representations for 3D shape generation tasks.
- Engineering conceptual design exploration
- Generating structurally plausible candidate geometric models from coarse block-outs during exploratory design phases.
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 structures rapidly, producing shapes within 2 to 4 seconds.
- Trained on a diverse collection of over 10 million shapes across 19 distinct 3D repositories.
- Uses compact wavelet-tree latent encodings to preserve fine details from TSDF representations at 256³ resolution.
- Demonstrates strong alignment to ground truth on standard benchmarks, achieving an IoU of 0.8285 and Chamfer Distance of 0.0020 on Google Scanned Objects.
Watch out for
- The low resolution of the 16³ input voxel grid can degrade output quality or cause ambiguous outputs.
- Shapes may occasionally be geometrically or structurally implausible, particularly if input conditioning is vague.
- Cannot be used for real-world load-bearing components or safety-critical structural applications as soundness is not guaranteed.
- Performance may degrade on domain-specific geometries or object categories that are underrepresented across the training sources.
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 VX16 1B commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What input format does WaLa VX16 1B expect?
The model expects a 16³ resolution voxel grid, which is passed through a ResNet convolutional encoder to generate the conditioning latent vector.
How fast is inference with this model?
The model generates 3D shapes in approximately 2 to 4 seconds.
Can this model be used to design load-bearing physical parts?
No. The model is intended strictly for research purposes and its outputs must not be relied upon for structural soundness or physical load-bearing applications.
How was the model evaluated?
It was tested on the Google Scanned Objects dataset (achieving an IoU of 0.8285, LFD of 1544.67, and CD of 0.0020) and on a validation subset of the training corpus (achieving an IoU of 0.75739, LFD of 1874.41, and CD of 0.0020).
How to cite it
Aditya Sanghi, Aliasghar Khani, Pradyumna Reddy, Arianna Rampini, Derek Cheung, Kamal Rahimi Malekshan, Kanika Madan, Hooman Shayani. Wavelet Latent Diffusion (Wala): Billion-Parameter 3D Generative Model with Compact Wavelet Encodings. arXiv:2411.08017, 2024.
Written from WaLa VX16 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.