Distributed CAD - Model Information

WaLa

Wala is stands for Wavelet Latent Diffusion. It was trained by compressing 3d model inputs from SDFs (signed distance functions) to wavelet trees via wavelet decomposition The wavelet representations are passed through a vector-quantized variational autoencoder (VQ-VAE) to generate a latent encoding, further compressing the input. This latent represntation is then passed through a DDPM (Diffusion Model) to generate a wavelet structure. This undergoes an inverse wavelet transform to generate the SDF, which is then rendered using a technique like marching cubes or poisson surface reconstruction.

This paper was created by researchers at Autodesk. More information can be found here.

Novel Architecture

The novel architecture has not yet been implemented. The data preprocessing and architecture implementation are still under research and development. It will be a combination of WaLa and PolyGen.

If you would like more information, feel free to contact me!

PolyGen

PolyGen is a transformer based architecture which generates very CAD-like models with smooth edges, flat faces, symmetry, etc... That being said, it requires specific prompts because the dataset(s) it was trained on do not encompass the vast assortment of data that WaLa had access to.

This paper was created by researchers at Autodesk. More information can be found here.