Open Collections

Wu, Yuhao

University of British Columbia

Deep generative models such as Variational Autoencoders (VAEs), Generative Ad- versarial Networks (GANs), and diffusion models have demonstrated their efficacy in generating 2D images and 3D meshes. However, interpreting the learned latent space poses a significant challenge. The current literature mainly focuses on un- supervised methods, which exhibit two primary limitations: firstly, the inability to control the meaning of each latent variable, and secondly, the occurrence of multiple latent variables possessing overlapping meanings. Moreover, it has been shown that fully disentangling the latent space using only unsupervised methods is theoretically infeasible. In this work, we introduce a method for latent space disen- tanglement on 3D meshes. Our method comprises two components: a feature func- tion for predicting 3D mesh features, and a regular generative model. We employ the derivative of the feature function as part of the loss function to promote dis- entanglement. Experimental results demonstrate that our disentanglement method effectively addresses the limitations mentioned above without compromising the accuracy of the reconstruction. Additionally, given its model-agnostic nature, our method exhibits generality across different generative models and tasks.

Text

2023-10-23

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/86287

Computer Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/