Multi-Modal Foundation · ADSKAILab
WaLa SV 1B
A 1-billion-parameter model generating detailed 3D shapes from single-view images using wavelet latent representations.
What WaLa SV 1B is
WaLa SV 1B is a 3D generative model developed by researchers at Autodesk AI Lab (ADSKAILab). It generates detailed three-dimensional geometry from a single 2D image input. The system addresses the computational challenge of scaling high-resolution 3D generation by encoding 3D shapes into compact wavelet-tree representations, enabling scalable generative modelling without excessive computational overhead.
Trained on an aggregation of over 10 million publicly available 3D shapes across diverse engineering and object datasets, WaLa SV 1B produces structural geometry in two to four seconds. The model is intended for academic research and scientific exploration into 3D generative frameworks across engineering and materials science contexts.
How it works
WaLa SV 1B uses a modified U-ViT (U-shaped Vision Transformer) architecture containing approximately 956 million parameters. During preprocessing, 3D training shapes are converted into truncated signed distance functions (TSDFs) at a resolution of 256³ and subsequently decomposed using a discrete wavelet transform into compact wavelet-tree latent encodings. This decomposition allows the network to capture both coarse geometric forms and fine structural details efficiently.
For single-view generation, an input 2D image is processed by a DINOv2 vision encoder to extract feature representations. These extracted visual features serve as conditioning latent vectors that guide the modified U-ViT diffusion framework to generate the corresponding 3D wavelet representation, which can then be converted into the final 3D shape.
Key facts
- Model parameters
- Approximately 956 million
- Architecture
- Modified U-ViT with DINOv2 conditioning encoder
- Latent representation
- Discrete wavelet transform of 256³ truncated signed distance functions (TSDF)
- Training data size
- Over 10 million 3D shapes across 19 public datasets
- Inference time
- 2 to 4 seconds
- Training hardware
- NVIDIA H100 GPUs
What you can use it for
- Single-View 3D Shape Reconstruction
- Generating complete 3D volumetric surfaces from single 2D input images for visual exploration and geometric analysis.
- Engineering Shape Conceptualisation
- Rapidly synthesising 3D object candidates from concept sketches or photographic views across diverse engineering domains.
- Generative Framework Research
- Evaluating wavelet-based latent representations and diffusion models on high-resolution 3D generative benchmark tasks.
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
- Fast inference, capable of producing detailed 3D shape outputs within 2 to 4 seconds.
- Trained on an extensive repository of over 10 million 3D shapes spanning 19 public datasets, including engineering datasets like ABC and Fusion 360.
- Wavelet-based compact latent encoding enables the effective capture of both fine details and coarse structures at 256³ TSDF resolution.
Watch out for
- Generated geometry cannot be relied upon for structural soundness or real-world load-bearing applications.
- Output quality degrades when provided with ambiguous, unclear, or low-quality input images.
- Performance may decline when generating object classes or styles that are underrepresented in the 19 public training collections.
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 WaLa SV 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 SV 1B accept?
It accepts a single-view 2D image, which is processed through a DINOv2 visual encoder to condition the 3D generation process.
How are 3D shapes internally represented in this model?
Shapes are preprocessed into 256³ truncated signed distance functions (TSDFs) and transformed into compact wavelet-tree latent encodings via discrete wavelet transform.
What datasets was the model trained on?
It was trained on more than 10 million shapes from 19 public 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.
Can generated meshes be used directly for structural engineering parts?
No. The model may occasionally generate implausible shapes, and even shapes that appear theoretically plausible are not verified for real-world load-bearing capacity or structural soundness.
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 [cs.CV], 2024.
Written from WaLa SV 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.