Reproduce “Maximum Spatial Perturbation Consistency for Unpaired Image-to-Image Translation” 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

Yanwu Xu, Shaoan Xie, Wenhao Wu, Kun Zhang, Mingming Gong, Kayhan Batmanghelich

Researchers from the University of Pittsburgh, Carnegie Mellon University, The University of Melbourne, and Baidu Inc. developed Maximum Spatial Perturbation Consistency (MSPC), a regularization technique for unpaired image-to-image translation. MSPC uses a learned, adaptive spatial perturbation function to dynamically challenge the translation model and align spatial content distributions, leading to improved robustness and superior performance across various benchmarks, including a new front-to-profile face translation task.

Key insights

  • Learning dynamic, image-dependent maximal spatial perturbations is more effective than fixed transformations for building I2I models that generalize beyond predefined patterns.
  • Enforcing commutativity between the translator and an adversarially optimized spatial perturbation function provides a powerful consistency regularization that improves content preservation and robustness.
  • An adversarially trained spatial perturbation function can simultaneously challenge the generator for robustness and actively align content distributions across domains, reducing spatial discrepancies.

Summary by alphaXiv.

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