Reproduce “AdaWCT: Adaptive Whitening and Coloring Style Injection” 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

Antoine Dufour, Yohan Poirier-Ginter, Alexandre Lessard, Ryan Smith, Michael Lockyer, Jean-Francois Lalonde

ADAWCT generalizes Adaptive Instance Normalization (AdaIN) by integrating Whitening and Coloring Transformation (WCT) principles for style injection in GANs. This method consistently improves image quality, achieving a 20% reduction in Fréchet Inception Distance (FID) over AdaIN in the StarGANv2 architecture, while also enhancing style-content disentanglement and reducing visual artifacts.

Key insights

  • Integrating higher-order statistical correlations via WCT principles into GANs' style injection mechanism can substantially improve generated image quality and realism compared to methods relying solely on first and second-order statistics.
  • Careful numerical stabilization and pre-processing steps are essential for effectively adapting WCT's matrix operations into deep learning architectures.
  • Group-wise processing provides an effective strategy to manage the parameter explosion associated with full covariance matrix computations, enabling a practical trade-off between computational cost and performance gains.

Summary by alphaXiv.

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