The GAN is dead; long live the GAN! A Modern GAN Baseline

arxiv:2501.05441

Original source · Paper JSON

Self-supervised learning and vision · 2025-01-09

Cites (1) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4
5. Part 5
6. Part 6
7. Part 7

Contents

• The GAN is dead; long live the GAN! A Modern Baseline GAN
• Abstract
• 1 Introduction
• 2 Serving Two Masters: Stability and Diversity with RpGAN + R 1 + R 2 +R_{1}+R_{2}
• 2.1 Traditional GAN
• 2.2 Relativistic f f -GAN
• 2.3 Training Dynamics of RpGAN
• Zero-centered gradient penalties.
• Discussion.
• 2.4 A Practical Demonstration
• 3 A Roadmap to a New Baseline — R3GAN
• Minimum baseline (Config B).
• Well-behaved loss function (Config C).
• General network modernization (Config D).
• Neural network architecture details.
• Bottleneck modernization (Config E).
• 4 Experiments Details
• 4.1 Roadmap Insights on FFHQ-256 Karras et al. [2019]
• 4.2 Mode Recovery — StackedMNIST Metz et al. [2016]
• 4.3 FID — FFHQ-256 Karras et al. [2019] (Optimized)
• 4.4 FID — FFHQ-64 Karras et al. [2022]
• 4.5 FID — CIFAR-10 Krizhevsky et al. [2009]
• 4.6 FID — ImageNet-32 Chrabaszcz et al. [2017]
• 4.7 FID — ImageNet-64 Chrabaszcz et al. [2017]
• 4.8 Recall
• 5 Discussion and Limitations
• 6 Conclusion
• Acknowledgements.
• Appendix A Local convergence
• Appendix B DiracRpGAN: A demonstration of non-convergence
• Summary.
• Definition B.1.
• Lemma B.2.
• Lemma B.3.
• Lemma B.4.
• Appendix C General Convergence Results
• Summary.
• C.1 Assumptions
• Assumption I.
• Assumption II.
• Assumption III.
• C.2 Convergence
• Theorem.
• Lemma C.2.1.
• Lemma C.2.2.
• Lemma C.2.3.
• Lemma C.2.4.
• Lemma C.2.5.
• Lemma C.2.6.
• Lemma C.2.7.
• Appendix D Hyperparameters, training configurations, and compute
• Training schedule.
• Dataset augmentation.
• Network capacity.
• Mixed precision training.
• Class conditioning.
• Stacked MNIST.
• Compute resources.
• Appendix E Negative Results and Future Work
• Appendix F Qualitative Results
• Appendix G Training Curves
• NeurIPS Paper Checklist
• Instructions for reporting errors