Reproduce “Generative Modeling using the Sliced Wasserstein Distance” 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
Ishan Deshpande, Ziyu Zhang, Alexander Schwing
The paper introduces the Sliced Wasserstein Generator (SWG), a method for generative modeling that replaces the unstable adversarial training of GANs with direct optimization of the Sliced Wasserstein Distance. This approach achieves improved training stability and robustness across various architectures and datasets, generating high-quality images on CIFAR-10, LSUN bedrooms, and CelebA while providing an interpretable loss metric.
Key insights
- While high-dimensional Wasserstein distance is complex, its 1D counterpart has a closed-form solution via sorting, which can be extended to high dimensions by averaging over random projections (Sliced Wasserstein Distance).
- Replacing the two-player adversarial game with a single-objective minimization of SWD transforms a notoriously unstable training problem into a more stable and direct optimization.
- The SWD itself serves as an interpretable metric that correlates well with the visual quality and diversity of generated samples, offering a clearer indication of training progress.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.