Reproduce “Bridging the Gap Between $f$-GANs and Wasserstein GANs” 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
Jiaming Song, Stefano Ermon
A new framework unifies f-GANs and Wasserstein GANs by introducing importance weights, enabling the proposed KL-Wasserstein GAN (KL-WGAN) to achieve the training stability and performance of WGANs alongside the density ratio interpretability of f-GANs. KL-WGAN demonstrates improved generation quality on image datasets like CIFAR10 and CelebA, and better distribution modeling on synthetic and UCI benchmarks compared to standard WGANs.
Key insights
- Both f-GANs and Wasserstein GANs can be viewed as specific instances within a generalized GAN framework by manipulating an importance weight variable.
- A new class of f-Wasserstein GANs can theoretically interpolate between f-divergences and the Wasserstein distance.
- For KL-divergence, the optimal importance weight function has a closed-form solution, enabling practical implementation without nested optimization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.