Research map / Learning theory and algorithms / Prediction and decision losses
How does Lipschitz Regularization Influence GAN Training?
Yipeng Qin, Niloy Mitra, Peter Wonka
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 |
|---|---|---|---|
| shrink domain with domain scaling alpha=1e-9 at k_SN=50 (ablation) | NS-GAN-SN · CelebA | FID: 242.64 -> 7.05 | Improved |
| shrink domain with domain scaling alpha=1e-9 at k_SN=50 (ablation) | COS-GAN-SN · CelebA | FID: 318.53 -> 6.31 | Improved |
| shrink domain with alpha=1e-25 at k_SN=50 instead of no scaling (ablation) | EXP-GAN-SN · CIFAR10 | FID: 401.24 -> 21.91 | Improved |
| use alpha=1e-25 domain scaling so loss functions match WGAN-SN (ablation) | NS-GAN-SN · CIFAR10 | FID: 23.36 -> 21.92 | Improved |
| lower k_SN from 1.0 to 0.2 (ablation) | NS-GAN-SN · CelebA | FID: 6.11 -> 18.59 | Hurt |
| raise k_SN from 1.0 to 50.0 (ablation) | COS-GAN-SN · CIFAR10 | FID: 15.88 -> 370.13 | Hurt |
| raise k_SN from 1.0 to 50.0 (ablation) | WGAN-SN · CIFAR10 | FID: 21.75 -> 23.36 | No clear effect |
| use WGAN loss instead of NS loss at k_SN=1.0 | WGAN-SN · CIFAR10 | FID: 15.81 -> 21.75 | Hurt |
| use non-standard cosine loss instead of NS loss at k_SN=1.0 | COS-GAN-SN · CelebA | FID: 6.11 -> 5.20 | No clear effect |
| use zero-centered LS-GAN# instead of LS-GAN with domain scaling | LS-GAN#-SN · CelebA (k_SN=50, alpha=1e-11) | FID: 135.17 -> 6.66 | Improved |
| reduce k_SN from 5.0 to 0.25 restricts domain of NS-GAN loss (ablation) | NS-GAN-SN · CelebA domain | domain width of discriminator output: [-8.130, 126.501] -> [-0.006, 0.006] | Improved |
| enlarge domain with alpha=1e5 at k_SN=1.0 (ablation) | LS-GAN-SN · MNIST | FID: 23.37 -> 260.05 | Hurt |
Closest papers by meaning
- Adversarial Lipschitz Regularization (ICLR 2020 · similarity 0.92)
- Loss-Sensitive Generative Adversarial Networks on Lipschitz Densities (IJCV 2019 · similarity 0.91)
- On the regularization of Wasserstein GANs (ICLR 2018 · similarity 0.91)
- Lipschitz Generative Adversarial Nets (ICML 2019 · similarity 0.90)
- Some Fundamental Aspects about Lipschitz Continuity of Neural Networks (ICLR 2024 · similarity 0.87)
- CHAIN: Enhancing Generalization in Data-Efficient GANs via lipsCHitz continuity constrAIned Normalization (CVPR 2024 · similarity 0.86)
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.
