Multi-Modal Foundation · ADSKAILab
WaLa DM6 1B
Generates 3D geometric shapes from six-view depth maps for engineering and materials research.
What WaLa DM6 1B is
WaLa DM6 1B is a 3D generative model developed by ADSKAILab as part of the Wavelet Latent Diffusion framework. It addresses the problem of reconstructing high-coverage, structurally detailed 3D geometries directly from multi-view surface observations.
The model conditions its generative process on six depth maps captured from pre-selected viewpoints around an object, producing completed 3D shapes within approximately four seconds. It bridges the gap between discrete sensor readings and dense 3D representations across engineering and materials science domains.
Trained on an aggregated collection of over 10 million publicly available 3D shapes across 19 sub-datasets, the system scales 3D generation to near one-billion parameters while retaining fine geometric features through a compact wavelet representation.
How it works
WaLa DM6 1B employs a modified U-ViT architecture operating on a compact wavelet-tree latent space. Training shapes are initially converted into truncated signed distance functions (TSDFs) at a resolution of 256³ voxels, which are subsequently decomposed using a discrete wavelet transform into multi-frequency wavelet-tree representations.
For conditional generation, each of the six input depth maps is passed through an individual DINO v2 encoder to generate a sequence of latent vectors. These multi-view feature vectors are concatenated into a joint conditioning representation that guides the latent diffusion network during 3D synthesis.
Key facts
- Architecture
- Modified U-ViT with DINO v2 depth encoder
- Parameter count
- Approximately 956 million
- Conditioning inputs
- Six-view depth maps
- Internal representation
- 256³ truncated signed distance function (TSDF) decomposed via discrete wavelet transform
- Generation time
- Approximately 4 seconds
- Training data
- Over 10 million shapes from 19 public datasets (including ShapeNet, ABC, Objaverse, and Fusion 360)
- Hardware
- NVIDIA H100 GPUs
What you can use it for
- Multi-View Depth-to-3D Synthesis
- Reconstructing complete 3D digital geometries from six calibrated depth camera views or simulated depth projections.
- Engineering Shape Synthesis
- Generating high-resolution geometric proposals and prototypes for engineering research using dense CAD and mechanical training priors.
- Geometric Modelling Research
- Evaluating wavelet-based latent diffusion representations against traditional voxel or point cloud generative approaches.
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
- Produces completed 3D geometric outputs in approximately four seconds.
- Achieves high surface accuracy, obtaining an Intersection over Union (IoU) of 0.91245 and Chamfer Distance of 0.00125 on the Google Scanned Objects benchmark.
- Encodes both fine structural details and overall shape plausibility by combining DINO v2 visual latents with discrete wavelet transforms.
- Trained on a broad variety of CAD, organic, and synthetic shapes across 19 major public datasets.
Watch out for
- Reconstruction fidelity degrades significantly if the provided depth maps are inaccurate, noisy, or ambiguous.
- Performance may drop when generating categories or styles poorly represented in the 10-million-shape training corpus.
- Generated outputs must not be used for load-bearing components or applications requiring real-world structural soundness guarantees.
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 DM6 1B commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What inputs does the model require?
It requires six depth maps captured from pre-selected viewpoints that collectively provide comprehensive visual coverage around the object.
How are the depth maps processed?
Each depth map is processed individually by a DINO v2 encoder into a sequence of latent vectors, which are then concatenated to condition the diffusion model.
What spatial representation is used for the generated 3D output?
The model generates shapes represented as truncated signed distance functions at 256³ resolution, decomposed using discrete wavelet transforms.
How long does shape generation take?
The model generates a completed 3D shape within approximately four seconds.
How to cite it
Aditya Sanghi, Aliasghar Khani, Pradyumna Reddy, Arianna Rampini, Derek Cheung, Kamal Rahimi Malekshan, Kanika Madan, and Hooman Shayani. Wavelet Latent Diffusion (Wala): Billion-Parameter 3D Generative Model with Compact Wavelet Encodings. arXiv:2411.08017, 2024.
Written from WaLa DM6 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.