Reproduce “Scaling up GANs for Text-to-Image Synthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, Taesung Park
GigaGAN demonstrates that Generative Adversarial Networks can be scaled to one billion parameters for open-world text-to-image synthesis, achieving competitive image quality (FID 9.09 on COCO2014) while generating images orders of magnitude faster (0.13 seconds for 512px) than iterative diffusion models. This work also showcases the continued utility of GANs' latent space controllability for image editing at scale.
Key insights
- GANs can be successfully scaled to billion-parameter levels and trained on vast, diverse datasets like LAION without inherent instability, challenging prior assumptions about their scalability.
- Careful architectural modifications to StyleGAN (including adaptive kernels, L2 attention, and a Multi-Scale Input, Multi-Scale Output discriminator) and novel loss functions (matching-aware and CLIP) are critical for stabilizing large-scale GAN training.
- GANs retain their advantage in inference speed and latent space controllability even at open-world text-to-image scale, offering a complementary paradigm to iterative diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.