Thanks to the org team: @junyanz.bsky.social @lingjieliu.bsky.social Deqing Sun, Lu Jiang, Fitsum Reda, and Krishna Kumar Singh!

Thanks, but I just twiddle my thumbs - it's all Nick and Aaron : )

arXiv: arxiv.org/abs/2501.05441
HuggingFace: huggingface.co/papers/2501....
OpenReview: openreview.net/forum?id=Ort...

We prioritize simplicity and performance over functionality. As a minimal baseline, our model does only basic image generation, lacking many features required for downstream tasks. Think of it as DCGAN in 2025 rather than something feature-rich like StyleGAN. We hope this helps further GAN research!

Given the well-behaved loss, we move away from the 2015-ish architecture in StyleGAN and implement G and D with a minimalist yet modern architecture---a simplified ConvNeXt. With the two components combined, we obtain a simple GAN baseline that is stable to train and surpasses StyleGAN performance.

To further GAN research, we first improve the GAN loss to alleviate mode dropping and non-convergence. This makes GAN optimization sufficiently easy that we can now discard existing GAN tricks w/o training failure. The dependence on outdated GAN-specific architectures is also eliminated.

GANs are often criticized for their training instability, and it is often believed that GANs cannot work w/o many engineering tricks. They use outdated network architectures without modern backbone advances. These supposed weaknesses resulted in the abandonment of GAN research in favor of diffusion.

Hey that's us! Let me know if anyone has any questions : )