Research map / Learning theory and algorithms / Prediction and decision losses
Towards a Better Global Loss Landscape of GANs
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 |
|---|---|---|---|
| Replace JS-GAN loss with relativistic pairing loss | Standard CNN GAN · CIFAR-10 | FID: 49.13 -> 33.79 | Helped |
| Replace JS-GAN loss with relativistic pairing loss | Standard CNN GAN · STL-10 | FID: 50.38 -> 52.54 | Hurt |
| Replace JS-GAN loss with relativistic pairing loss while using spectral normalization | ResNet GAN · CIFAR-10 | FID: 20.13 -> 19.31 | Helped |
| Replace JS-GAN loss with relativistic pairing loss while using spectral normalization | ResNet GAN · STL-10 | FID: 36.33 -> 34.77 | Helped |
| Use relativistic pairing loss instead of JS-GAN loss with spectral normalization | ResNet with generator and discriminator channels halved · CIFAR-10 | FID: 23.29 -> 21.78 | Helped |
| Use relativistic pairing loss instead of JS-GAN loss with spectral normalization | ResNet with generator and discriminator channels reduced to one quarter · CIFAR-10 | FID: 45.20 -> 31.26 | Helped |
| Use relativistic pairing loss instead of JS-GAN loss with spectral normalization | ResNet with generator and discriminator channels halved · STL-10 | FID: 44.39 -> 42.18 | Helped |
| Use relativistic pairing loss instead of JS-GAN loss with spectral normalization | ResNet with generator and discriminator channels reduced to one quarter · STL-10 | FID: 53.57 -> 52.09 | Helped |
| Increase network width (ablation) | JS-GAN and RS-GAN · Two-cluster synthetic data | convergence speed | Helped |
| Use a narrow network (ablation) | RS-GAN instead of JS-GAN · Two-cluster synthetic data | mode recovery within 5k iterations: all five runs; two of the five runs | Helped |
| Use RS-GAN instead of JS-GAN | Two-cluster synthetic data · Two-cluster synthetic data | iterations to escape mode collapse: 800 iterations; about 3 times faster | Helped |
| Use spectral normalization (ablation) | JS-GAN · CIFAR-10, Standard CNN | FID: 49.13 -> 28.07 | Helped |
Closest papers by meaning
- The GAN is dead; long live the GAN! A Modern GAN Baseline (NeurIPS 2024 · similarity 0.91)
- A Domain Agnostic Measure for Monitoring and Evaluating GANs (NeurIPS 2019 · similarity 0.89)
- A Classification-Based Study of Covariate Shift in GAN Distributions (ICML 2018 · similarity 0.88)
- A Neural Tangent Kernel Perspective of GANs (ICML 2022 · similarity 0.88)
- On the Limitations of First-Order Approximation in GAN Dynamics (ICML 2018 · similarity 0.88)
- 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.