Research map / Image generation / GANs
Wasserstein Divergence for GANs
Jiqing Wu, Zhiwu Huang, Janine Thoma, Dinesh Acharya, Luc Van Gool
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace WGAN-GP gradient penalty with Wasserstein divergence objective | WGAN-GP · LSUN (standard training) | FID: 20.3 -> 15.9 | Improved |
| use Wasserstein divergence objective instead of CTGAN | CTGAN · CIFAR-10 (standard training) | FID: 18.6 -> 18.1 | Improved |
| use Wasserstein divergence objective instead of WGAN-GP | WGAN-GP · CelebA 64x64 (ResNet) | FID: 18.4 -> 15.2 | Improved |
| use Wasserstein divergence instead of WGAN-GP when batch norm is removed from ResNet | WGAN-GP · CelebA 64x64 (ResNet without BN) | FID: 20.3 -> 18.6 | Improved |
| use Wasserstein divergence instead of WGAN-GP with ConvNet | WGAN-GP · CelebA 64x64 (ConvNet) | FID: 21.2 -> 17.5 | Improved |
| use Wasserstein divergence instead of WGAN-GP with ConvNet without BN | WGAN-GP · CelebA 64x64 (ConvNet without BN) | FID: 24.6 -> 21.5 | Improved |
| replace the PGGAN objective with the W-div objective (PGGAN-div) at 64x64 | PGGAN · CelebA 64x64 (progressive) | FID: 16.3 -> 16.0 | Improved |
| replace the PGGAN objective with the W-div objective at 128x128 | PGGAN · LSUN 128x128 (progressive) | FID: 15.4 -> 15.5 | No clear effect |
| replace the PGGAN objective with the W-div objective at 256x256 | PGGAN · LSUN 256x256 (progressive) | FID: 15.1 -> 14.9 | Improved |
| increase power p from 2 to the optimum p=6 in the divergence gradient term (ablation) | WGAN-div · CelebA 64x64 | FID | Improved |
| vary coefficient k with p fixed at 6 (ablation) | WGAN-div · CelebA 64x64 | FID: FID fluctuating mildly around 16 | No clear effect |
| change sampling strategy of P_u from real-fake interpolation (1) to sampling real points only (6) (ablation) | WGAN-div · CelebA 64x64 | FID: 15.2 -> 14.9 | Improved |
Closest papers by meaning
- Towards Generalized Implementation of Wasserstein Distance in GANs (AAAI 2021 · similarity 0.94)
- A Two-Step Computation of the Exact GAN Wasserstein Distance (ICML 2018 · similarity 0.93)
- Some Theoretical Insights into Wasserstein GANs (JMLR 2021 · similarity 0.93)
- Improving the Improved Training of Wasserstein GANs: A Consistency Term and Its Dual Effect (ICLR 2018 · similarity 0.93)
- Banach Wasserstein GAN (NeurIPS 2018 · similarity 0.92)
- Improved Training of Wasserstein GANs (NeurIPS 2017 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
