Multi-Modal Foundation · ADSKAILab
WaLa DM1 1B
Generates detailed 3D shapes from a single-view depth map in four seconds using wavelet latent diffusion.
What WaLa DM1 1B is
WaLa-DM1-1B is a 3D generative model developed by Aditya Sanghi, Aliasghar Khani, Chinthala Pradyumna Reddy, Arianna Rampini, Derek Cheung, Kamal Rahimi Malekshan, Kanika Madan, and Hooman Shayani at ADSKAILab. The system addresses the challenge of reconstructing detailed, structurally plausible 3D geometries from minimal visual input, specifically a single-view depth map.
Trained on more than 10 million publicly available 3D shapes, the model generates complex geometries in approximately 4 seconds. It is designed for scientific research and development, allowing researchers in engineering and materials science to explore minimal-input 3D shape synthesis.
How it works
The model is built on a modified U-ViT architecture operating over a compact, wavelet-based latent encoding. Each training shape is preprocessed into a truncated signed distance function (TSDF) at 256³ resolution, which is then decomposed via a discrete wavelet transform into a wavelet-tree representation to capture both coarse and fine spatial features.
Conditioning is derived from a single depth map processed through a DINOv2 encoder. The resulting latent sequence forms the conditioning vectors for the generative diffusion process. Training was executed across NVIDIA H100 GPUs using an aggregate of 19 sub-datasets, including ModelNet, ShapeNet, ABC, Fusion 360, Objaverse, and subsets of ObjaverseXL.
Key facts
- Architecture
- Modified U-ViT with DINOv2 conditioning encoder
- Parameter count
- Approximately 956 million
- Input format
- Single-view depth map
- Representation
- Wavelet-tree latent derived from 256³ TSDF
- Inference time
- Approximately 4 seconds
- Hardware used
- NVIDIA H100 GPUs
What you can use it for
- Minimal-Input 3D Reconstruction
- Reconstruct full 3D geometries from single-perspective depth captures for experimental visualization.
- Generative Shape Prototyping
- Rapidly synthesize candidate 3D structures from synthetic or captured depth profiles in research exploration.
- Wavelet-Tree Generative Research
- Study the theoretical behaviour and fidelity of multiscale wavelet-latent representations in 3D diffusion models.
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
- Generates complete 3D structures from only a single depth map.
- Achieves rapid shape generation in roughly 4 seconds.
- Employs a wavelet-based representation to preserve fine geometric details alongside coarse structure.
- Trained across a diverse dataset of over 10 million 3D shapes spanning 19 public repositories.
Watch out for
- Performance degrades when provided with ambiguous, inaccurate, or low-quality input depth maps.
- Generated meshes may lack real-world structural soundness and must not be used for load-bearing applications.
- Reconstruction fidelity can vary across object categories underrepresented in the training data.
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 DM1 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-DM1-1B require?
It takes a single-view depth map as conditioning input, which is encoded using DINOv2.
What 3D representation does the model produce?
The model works over a wavelet-tree representation derived from a 256³ truncated signed distance function (TSDF).
How well does the model perform on standard benchmarks?
On the Google Scanned Objects benchmark, it achieves an Intersection over Union (IoU) of 0.6927, a Chamfer Distance of 0.01301, and a Light Field Distance of 2172.52.
Can generated models be used directly for structural engineering components?
No. The model occasionally produces implausible geometries, and even structurally plausible outputs should not be relied upon for load-bearing physical applications.
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 DM1 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.