# Scaling up GANs for Text-to-Image Synthesis

**Authors:** Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, Taesung Park  
**arXiv:** [2303.05511](https://arxiv.org/abs/2303.05511)  
**Date:** 2023-03-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.05511v1/

## Key points

- Text-to-image synthesis has revolutionized the field of generative image models
- Auto-regressive and diffusion models like DALL-E 2 have become the new standard for large-scale generative models
- GANs can be scaled up to benefit from large datasets like LAION with a new approach
- A team of researchers led by Minguk Kang developed GigaGAN, a new GAN architecture that far exceeds the capacity limit of StyleGAN architecture
- GigaGAN offers three major advantages over other architectures:
- It is orders of magnitude faster at inference time
- It can synthesize high-resolution images quickly
- It supports various latent space editing applications such as latent interpolation, style mixing, and vector arithmetic operations
- GigaGAN outperforms existing methods on several metrics including FID and Inception Score when trained on both small and large datasets such as LSUN and LAION
- GigaGAN is effective in generating realistic images from textual descriptions using COCO-Stuff and CUB datasets.
- Scaling up GANs is not only possible but also highly effective with the right approach.

## Summary

The recent success of text-to-image synthesis has sparked a revolution in the field of generative image models. While GANs were once the go-to choice for this task, auto-regressive and diffusion models like DALL-E 2 have become the new standard for large-scale generative models overnight. This shift has raised an important question: can we scale up GANs to benefit from large datasets like LAION? The answer is yes, and it requires a new approach. Naively increasing the capacity of the StyleGAN architecture quickly becomes unstable, which is why a team of researchers led by Minguk Kang developed GigaGAN, a new GAN architecture that far exceeds this limit. GigaGAN offers three major advantages over other architectures. First, it is orders of magnitude faster at inference time, taking only 0.13 seconds to synthesize a 512px image. Second, it can synthesize high-resolution images quickly - for example, 16-megapixel pixels in just 3.66 seconds. Finally, GigaGAN supports various latent space editing applications such as latent interpolation, style mixing, and vector arithmetic operations. The team's experiments show that GigaGAN outperforms existing methods on several metrics including FID and Inception Score when trained on both small and large datasets such as LSUN and LAION. They also demonstrate its effectiveness in generating realistic images from textual descriptions using COCO-Stuff and CUB datasets. Overall, this research shows that scaling up GANs is not only possible but also highly effective with the right approach. By introducing GigaGAN as a viable option for text-to-image synthesis, this work opens up exciting possibilities for future research in generative image modeling.