Published in BMVC 2021

Learning Texture Generators for 3D Shape Collections
from Internet Photo Sets

rendering resutls

          Rendering results of synthesized textures conditioned on the shape. For each example, we vary the shape, shown for 4 different views, over           the columns (i.e., fixed z) and the texture over the rows (i.e., different z).


We present a method for decorating existing 3D shape collections by learning a tex- ture generator from internet photo collections. We condition the StyleGAN [15] texture generation by injecting multiview silhouettes of a 3D shape with SPADE-IN [23]. To bridge the inherent domain gap between the multiview silhouettes from the shape col- lection and the distribution of silhouettes in the photo collection, we employ a mixture of silhouettes from both collections for training. Furthermore, we do not assume each exemplar in the photo collection is viewed from more than one vantage point, and lever- age multiview discriminators to promote semantic view-consistency over the generated textures. We verify the efficacy of our design on three real-world 3D shape collections.


GAN, Texture generation

Paper and video

Trained model and code


Pieter Peers was supported in part by NSF grant IIS-1909028.