Research map / Image generation / GANs
Many Paths to Equilibrium: GANs Do Not Need to Decrease a Divergence At Every Step
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 |
|---|---|---|---|
| Use the non-saturating generator loss rather than the minimax loss | GAN · Synthetic 1-D manifold task with disjoint support | ability to fit the data distribution | Helped |
| Add a gradient penalty near the data manifold (DRAGAN-NS) (ablation) | Non-saturating GAN · Synthetic experiments | convergence stability | Helped |
| Add an interpolation-based gradient penalty (GAN-GP) (ablation) | Non-saturating GAN · Synthetic experiments | convergence stability | Helped |
| Add GAN-GP or DRAGAN-NS gradient penalty (ablation) | Non-saturating GAN · Color MNIST, CelebA, CIFAR-10 | robustness to hyperparameter changes | Helped |
| Add gradient penalties (ablation) | Non-saturating GAN · CIFAR-10 | Inception Score | Helped |
| Increase discriminator updates to 100 per generator update (ablation) | Non-saturating GAN · Synthetic experiments | distribution fit | Hurt |
| Add GAN-GP while increasing discriminator update count (ablation) | Non-saturating GAN · Synthetic experiments | scaling with discriminator updates | Helped |
| Reduce WGAN-GP to one discriminator update per generator update (ablation) | WGAN-GP · Color MNIST, CelebA, CIFAR-10 | sample quality | Hurt |
| Apply both GAN-GP and DRAGAN-NS penalties together (ablation) | Non-saturating GAN · Real data | model quality | No clear effect |
| Replace the non-saturating GAN with WGAN-GP | GAN training · CelebA | ability to learn the data distribution | Mixed |
| Increase the data-space dimension while keeping the learning rate fixed (ablation) | Non-saturating GAN · Synthetic Experiment 1 | fit strength | Hurt |
| Use an overcomplete generator with more latent directions than the data manifold requires (ablation) | Non-saturating GAN · Synthetic Experiment 2 | ability to capture the data distribution | No clear effect |
Closest papers by meaning
- Do GANs always have Nash equilibria? (ICML 2020 · similarity 0.92)
- Finding Mixed Nash Equilibria of Generative Adversarial Networks (ICML 2019 · similarity 0.92)
- On Characterizing GAN Convergence Through Proximal Duality Gap (ICML 2021 · similarity 0.91)
- A Convex Duality Framework for GANs (NeurIPS 2018 · similarity 0.90)
- Approximation and Convergence Properties of Generative Adversarial Learning (NeurIPS 2017 · similarity 0.90)
- Generative Adversarial Equilibrium Solvers (ICLR 2024 · similarity 0.89)
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.