Research map / Architectures and efficiency / RNNs and normalization
The GAN is dead; long live the GAN! A Modern GAN Baseline
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes R3GAN, a minimalist modern GAN baseline built on a regularized relativistic pairing GAN loss (RpGAN + R1 + R2) and a modernized ResNet/ConvNeXt-style backbone, showing competitive FID without the ad-hoc tricks of StyleGAN2.
Problem
GANs are widely considered hard to train and rely on many empirical tricks; the paper asks whether a principled loss and modern backbone can yield a simple GAN baseline that converges reliably and matches or beats strong GANs and diffusion models.
Approach
Derive a regularized relativistic pairing GAN loss (RpGAN) with both R1 and R2 gradient penalties that has local and global convergence guarantees, then strip StyleGAN2 of its tricks and rebuild the generator and discriminator as a modernized ResNet (1-3-1 bottlenecks, grouped convolutions, inverted bottleneck, fixup init, no normalization).
Key results
- On FFHQ-256 the final R3GAN reaches FID 7.05 versus 7.52 for the StyleGAN2 baseline (Config A) under the same ~25M-parameter budget.
- On StackedMNIST the model covers all 1000 modes with KL divergence 0.029, the best among the compared GANs and likelihood-based methods.
- The RpGAN+R1+R2 loss alone, without StyleGAN2 tricks, achieves FID 11.65 at Config C, and the modernized backbone improves it to 9.95 in Config D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add R2 gradient penalty to RpGAN + R1 (ablation) | RpGAN · StackedMNIST | training convergence | Helped |
| Use RpGAN instead of conventional GAN with R1 + R2 (ablation) | GAN with R1 + R2 · StackedMNIST | D_KL: 0.9270 to 0.0781 | Helped |
| Remove StyleGAN2 features and training tricks (ablation) | StyleGAN2 on FFHQ-256 · FFHQ-256 | FID: 7.516 -> 12.46 | Hurt |
| Add RpGAN loss and R2 gradient penalty (ablation) | Stripped StyleGAN2 · FFHQ-256 | FID: 12.46 -> 11.65 | Helped |
| Modernize G and D with ResNet blocks (ablation) | Config C · FFHQ-256 | FID: 11.65 -> 10.17 | Helped |
| Remove generator output skips (ablation) | Config D · FFHQ-256 | FID: 10.17 -> 9.950 | Helped |
| Add grouped convolutions and reduce bottleneck compression ratio to two (ablation) | Config D · FFHQ-256 | FID: 9.950 -> 7.507 | Helped |
| Invert stem and bottleneck widths (ablation) | Config E with grouped convolutions · FFHQ-256 | FID: 7.507 -> 7.045 | Helped |
| Use large-kernel depthwise convolutions with fewer channels instead of wider 3×3 convolutions (ablation) | Config D · FFHQ-256 | FID | Hurt |
| Swap Leaky ReLU for GELU (ablation) | GAN generator and discriminator · GAN image generation | FID | Hurt |
| Swap Leaky ReLU for PReLU (ablation) | GAN generator and discriminator · GAN image generation | FID | Helped |
| Add minibatch standard deviation (ablation) | StyleGAN · StackedMNIST | modes recovered: 857 -> 881 | Helped |
Surprising findings
- RpGAN with only R1 regularization diverges quickly regardless of the gamma value, so R2 is needed for stable training.
- Karras et al.'s conclusion that ResNet generators are harmful to StyleGAN2 does not hold once the ResNet is modernized and output skips are removed.
Limitations
- Not evaluated on higher-resolution ImageNet or large-scale text-to-image generation, so scalability is unverified.
- Omits adaptive normalization and multi-head self-attention, which prior work shows improve FID, and drops style injection that is useful for editing.
- Does not claim to beat every existing model on every dataset or task.
Experimental setup
Datasets: StackedMNIST FFHQ-256 CIFAR-10 ImageNet-32
Benchmarks: StackedMNIST FFHQ-256 CIFAR-10 ImageNet-32
Metrics: FID number of modes KL divergence
Baselines: StyleGAN2 StyleGAN-XL DCGAN WGAN-GP PresGAN DDGAN MEG PacGAN VAEBM Adv. DSM
Builds on: StyleGAN2 Relativistic GAN ConvNeXt Mescheder et al. local convergence analysis Fixup initialization
Closest papers by meaning
- Re-GAN: Data-Efficient GANs Training via Architectural Reconfiguration (CVPR 2023 · similarity 0.92)
- Towards a Better Global Loss Landscape of GANs (NeurIPS 2020 · similarity 0.91)
- Are GANs Created Equal? A Large-Scale Study (NeurIPS 2018 · similarity 0.91)
- A Large-Scale Study on Regularization and Normalization in GANs (ICML 2019 · similarity 0.91)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.89)
- SMaRt: Improving GANs with Score Matching Regularity (ICML 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.