Research map / Image generation / GANs
Improving Generalization and Stability of Generative Adversarial Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add zero-centered gradient penalty on interpolations between real and fake samples | GAN · 8 Gaussians | mode coverage and interpolation | Helped |
| Add zero-centered gradient penalty on interpolations between real and fake samples | GAN · MNIST | training stability and sample quality | Helped |
| Add zero-centered gradient penalty on interpolations between real and fake samples | GAN · ImageNet | Inception score | Helped |
| Use zero-centered gradient penalty on real/fake samples only instead of on interpolations (ablation) | GAN · 8 Gaussians | mode coverage and interpolation | Hurt |
| Use zero-centered gradient penalty on real/fake samples only instead of on interpolations (ablation) | GAN · ImageNet | Inception score | Hurt |
| Replace zero-centered gradient penalty on interpolations with one-centered gradient penalty (ablation) | GAN · scaled 8 Gaussians | mode coverage | Hurt |
| Replace zero-centered gradient penalty on interpolations with one-centered gradient penalty (ablation) | GAN · MNIST | training stability and sample quality | Hurt |
| Increase the number of discriminator updates per generator update to 5 (ablation) | GAN without zero-centered gradient penalty or with other penalties · MNIST | learning | Hurt |
| Increase the number of discriminator updates per generator update to 5 (ablation) | GAN with zero-centered gradient penalty · MNIST | learning | Helped |
| Increase the zero-centered gradient penalty weight (ablation) | GAN-0-GP · MNIST | diversity of generated samples | Helped |
| Set the zero-centered gradient penalty weight to 50 (ablation) | GAN-0-GP · MNIST | sample similarity | Hurt |
| Use Adam with large $\beta_1$ (e.g. 0.9) (ablation) | GANs with other gradient penalties · MNIST | learning | Hurt |
Closest papers by meaning
- On the Discrimination-Generalization Tradeoff in GANs (ICLR 2018 · similarity 0.94)
- Training Generative Adversarial Networks with Limited Data (NeurIPS 2020 · similarity 0.92)
- On Stabilizing Generative Adversarial Training With Noise (CVPR 2019 · similarity 0.92)
- Learning GANs and Ensembles Using Discrepancy (NeurIPS 2019 · similarity 0.92)
- Unrolled Generative Adversarial Networks (ICLR 2017 · similarity 0.91)
- Stabilizing Training of Generative Adversarial Networks through Regularization (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.