Research map / Image generation / GANs
Alleviation of Gradient Exploding in GANs: Fake Can Be Real
Song Tao, Jia Wang
Check 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 |
|---|---|---|---|
| Apply fake-as-real selection rather than real-sample zero-centered gradient penalty alone | Conventional NSGAN · CIFAR-10 (500k) | Inception score: 6.94\pm 0.03 -> 7.63\pm 0.05 | Improved |
| Apply fake-as-real selection rather than real-sample zero-centered gradient penalty alone | Conventional NSGAN · CIFAR-10 (500k) | FID: 16.66\pm 0.14 -> 12.80\pm 0.31 | Improved |
| Apply fake-as-real selection rather than real-sample zero-centered gradient penalty alone | Conventional NSGAN · CIFAR-100 (500k) | Inception score: 6.92\pm 0.08 -> 7.84\pm 0.04 | Improved |
| Apply fake-as-real selection rather than real-sample zero-centered gradient penalty alone | Conventional NSGAN · CIFAR-100 (500k) | FID: 22.28\pm 0.45 -> 17.69\pm 0.24 | Improved |
| Apply fake-as-real selection rather than real-sample zero-centered gradient penalty alone | ResNet NSGAN · ImageNet (600k) | Inception score: 10.66\pm 0.11 -> 11.44\pm 0.05 | Improved |
| Apply fake-as-real selection rather than real-sample zero-centered gradient penalty alone | ResNet NSGAN · ImageNet (600k) | FID: 44.57\pm 0.34 -> 39.69\pm 0.57 | Improved |
| Select fake samples randomly as real ones (f=1), instead of selecting low-discriminator-output samples (ablation) | FARGAN · CIFAR-10 | Generation performance | No clear effect |
| Select low-discriminator-output fake samples from a larger candidate set (f=8) rather than random fake samples (f=1) (ablation) | FARGAN · CIFAR-10 | Generation performance | Improved |
| Increase the candidate-pool multiplier f up to a sufficiently large value (e.g. f=8) (ablation) | FARGAN · CIFAR-10 | Generation performance | Improved |
| Increase the number of fake-as-real samples N0 appropriately (ablation) | FARGAN · CIFAR-10 | Generation performance | Improved |
| Set the number of fake-as-real samples to N0=48 (ablation) | FARGAN · CIFAR-10 | Generation performance | Hurt |
Closest papers by meaning
- Tempered Adversarial Networks (ICML 2018 · similarity 0.89)
- Smoothness and Stability in GANs (ICLR 2020 · similarity 0.88)
- FakeCLR: Exploring Contrastive Learning for Solving Latent Discontinuity in Data-Efficient GANs (ECCV 2022 · similarity 0.88)
- Training Generative Adversarial Networks by Solving Ordinary Differential Equations (NeurIPS 2020 · similarity 0.88)
- Top-k Training of GANs: Improving GAN Performance by Throwing Away Bad Samples (NeurIPS 2020 · similarity 0.87)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.87)
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.
