Reproduce “Non-Stationary Texture Synthesis by Adversarial Expansion” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, Hui Huang

A method called "adversarial expansion" is introduced to synthesize non-stationary textures by training a GAN to double the spatial extent of input texture blocks. This approach accurately reproduces and extends complex patterns like wood grain and peacock feathers, demonstrating scalability up to 32 times the original size.

Key insights

  • Residual blocks within the generator network are instrumental in synthesizing new large-scale structures, such as additional concentric rings in wood grain or branching veins in leaf textures.
  • The adversarial loss is a primary component for ensuring perceptual realism, while the L1 loss helps stabilize training and the VGG-based style loss contributes to finer local details.
  • The adversarial expansion approach demonstrates stability under repeated application, enabling coherent texture synthesis up to 32 times the original size.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.