Multi-Modal Foundation · ADSKAILab
WaLa SK 1B
A billion-parameter 3D generative model that synthesises detailed 3D shapes from 2D sketch inputs.
What WaLa SK 1B is
WaLa SK 1B (Wavelet Latent Diffusion Sketch 1B) is a 3D generative model designed to synthesise complex three-dimensional geometries directly from 2D sketch inputs. Developed by Autodesk AI Lab researchers, the model enables designer-driven shape synthesis from simple line drawings, producing high-detail meshes within two to four seconds.
The model addresses the challenge of recovering rich, plausible geometric structures from sparse 2D representations. By combining wavelet-based latent representations with diffusion modelling, it allows engineering and materials science researchers to rapidly explore 3D concepts from quick sketches.
How it works
WaLa SK 1B uses a modified U-ViT (U-shaped Vision Transformer) backbone operating on a compact wavelet-based latent space. Training 3D shapes are preprocessed into truncated signed distance functions (TSDF) at a resolution of 256³ and then decomposed using a discrete wavelet transform into wavelet-tree representations. This structure captures both coarse volumetric outlines and high-frequency geometric details.
The model was initially trained on over 10 million publicly available 3D shapes drawn from 19 sub-datasets using NVIDIA H100 GPUs. It was subsequently fine-tuned on synthetic sketch data generated via six distinct rendering methods, including Canny edge detection, Holistically-Nested Edge Detection (HED), Grease Pencil, and CLIPasso.
Key facts
- Architecture
- Modified U-ViT with compact wavelet latent encoding
- Parameter count
- Approximately 956 million
- Input format
- 2D sketch / line art
- Target representation
- Truncated signed distance function (TSDF) at 256³ resolution
- Inference speed
- 2–4 seconds per shape
- Pre-training dataset size
- Over 10 million 3D shapes across 19 public sub-datasets
- Hardware used
- NVIDIA H100 GPUs
What you can use it for
- Sketch-to-3D concept prototyping
- Translates hand-drawn sketches or digital line art into 3D geometric shapes within 2 to 4 seconds for early-stage design iteration.
- Exploration of 3D generative frameworks
- Provides a large-scale research baseline for assessing wavelet-tree latent diffusion architectures applied to spatial and physical shape generation.
- Line art shape reconstruction studies
- Supports investigation into how synthetic sketch rendering techniques, such as CLIPasso and Canny filters, influence 3D geometric inference.
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
- Rapid shape generation, outputting detailed 3D structures in 2 to 4 seconds.
- Captures both coarse and fine geometric features using a compact wavelet-tree TSDF representation.
- Pre-trained on a diverse foundation of over 10 million 3D shapes and fine-tuned across six synthetic sketch generation pipelines.
Watch out for
- Generated geometry cannot be relied upon for real-world structural soundness or load-bearing applications.
- Performance may degrade when evaluated on object categories or stylistic variations underrepresented in the training data.
- Ambiguous or poor-quality input sketches can cause the model to output geometrically implausible shapes.
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 SK 1B commercially?
Check the licence before commercial use. The other licence carries terms worth reading before you depend on it.
What kind of inputs does WaLa SK 1B require?
The model takes 2D sketches or line art as input, which it uses as conditioning to generate full 3D shapes.
How is 3D geometry represented inside the model?
Shapes are represented as truncated signed distance functions (TSDF) at 256³ resolution, decomposed into compact wavelet-tree structures via discrete wavelet transforms.
Can the generated shapes be used directly in engineering load-bearing components?
No. The model is intended strictly for research exploration and should not be used for creating load-bearing physical objects, as theoretical plausibility does not ensure real-world structural integrity.
How was the sketch fine-tuning dataset constructed?
Synthetic sketches were generated from 3D models using six different methods, including Canny edge detection, HED, Grease Pencil, and CLIPasso.
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 SK 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.