Multi-Modal Foundation · ADSKAILab
WaLa UN 1B
A billion-parameter 3D diffusion model generating detailed shapes unconditionally from noise for design exploration.
What WaLa UN 1B is
WaLa UN 1B is an unconditional 3D generative model developed by ADSKAILab. Built to address design-space exploration in engineering and materials science, it synthesises diverse, complex 3D structures directly from random noise without requiring text prompts, images, or external conditioning inputs.
The model forms part of the Wavelet Latent Diffusion (WaLa) framework. By operating on a compact wavelet-based latent representation, the model achieves high structural plausibility and fine geometric fidelity while generating complete 3D shapes in 2 to 4 seconds.
WaLa UN 1B is intended for academic research and the theoretical demonstration of generative 3D architectures across varied shape distributions.
How it works
WaLa UN 1B uses a modified U-ViT (U-shaped Vision Transformer) architecture comprising approximately 1.1 billion parameters. During preprocessing, training shapes are converted into truncated signed distance functions (TSDFs) at a resolution of 256³ and decomposed via a discrete wavelet transform into a compact wavelet-tree representation. The model then learns a diffusion process across this wavelet latent space.
Training was conducted on NVIDIA H100 GPUs using an aggregated corpus of more than 10 million publicly available 3D shapes drawn from 19 datasets, including ModelNet, ShapeNet, Thingi10K, ABC, Fusion 360, BuildingNet, Objaverse, and ObjaverseXL subsets. Generation proceeds purely from random noise without condition inputs.
Key facts
- Architecture
- Modified U-ViT (Wavelet Latent Diffusion)
- Parameter count
- 1.1 billion
- Input representation
- Random noise
- Output representation
- Wavelet-encoded truncated signed distance function (TSDF) at 256³ resolution
- Inference speed
- 2–4 seconds
- Training data
- Over 10 million 3D shapes from 19 public datasets
- Hardware used
- NVIDIA H100 GPUs
What you can use it for
- Unconditional Design-Space Exploration
- Generate diverse structural shapes from random initialisations to survey synthetic geometric variations in computational design.
- Generative 3D Framework Evaluation
- Benchmark wavelet-tree latent diffusion architectures against alternative 3D representations without conditioning constraints.
- Synthetic 3D Data Generation
- Synthesise high-resolution geometric forms to supplement downstream 3D machine learning tasks and shape-distribution studies.
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 complete 3D shapes with detailed geometry and complex structures in 2 to 4 seconds.
- Leverages a compact wavelet latent encoding that effectively captures both coarse and fine geometric features.
- Trained on a diverse collection of over 10 million shapes spanning 19 public datasets.
Watch out for
- Offers no direct control or prompt steering over the type, category, or attributes of the generated shapes.
- May generate implausible, unrealistic, or structurally unsound geometries that cannot be used for load-bearing physical applications.
- Performance and representation vary across object classes due to uneven distributions in the public training datasets.
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 UN 1B commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
Can I provide text or images to guide the generation in WaLa UN 1B?
No. WaLa UN 1B is strictly an unconditional model that generates 3D geometries solely from random noise inputs.
What geometric format does the model operate on?
The model operates on a wavelet-tree representation derived from a 256³ resolution truncated signed distance function (TSDF) decomposed with a discrete wavelet transform.
Can generated models be used directly for engineering manufacturing?
No. The model may generate implausible shapes, and outputs must not be relied upon for structural soundness or load-bearing physical objects.
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 UN 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.