# Large Scale GAN Training for High Fidelity Natural Image Synthesis

**Authors:** Andrew Brock, Jeff Donahue, Karen Simonyan  
**arXiv:** [1809.11096](https://arxiv.org/abs/1809.11096)  
**Date:** 2018-09-28  
**Summary page:** https://www.summarizepaper.com/en/paper/1809.11096v1/

## Key points

- Recent progress in generative image modeling for complex datasets like ImageNet
- Authors train Generative Adversarial Networks (GANs) at an unprecedented scale
- Investigation of instabilities associated with large-scale GAN training
- Truncating the latent space allows for fine control over sample fidelity and variety
- Development of BigGAN models, setting a new state-of-the-art in class-conditional image synthesis
- Impressive performance on ImageNet: Inception Score (IS) of 166.3 and Frechet Inception Distance (FID) of 9.6
- Significant improvements over previous best IS and FID scores
- Training GANs at larger scale with orthogonal regularization and truncation trick enables high-fidelity, diverse image generation from complex datasets like ImageNet
- Findings contribute to advancements in generative image modeling with implications for computer vision and artificial intelligence applications.

## Summary

In the field of generative image modeling, there has been recent progress in generating high-resolution and diverse samples from complex datasets like ImageNet. To address this challenge, the authors of this study train Generative Adversarial Networks (GANs) at an unprecedented scale and investigate the specific instabilities associated with such large-scale training. This trick allows for fine control over the trade-off between sample fidelity and variety by truncating the latent space. As a result of these modifications, the authors develop models known as BigGANs, which set a new state-of-the-art in class-conditional image synthesis. To evaluate their models' performance, they train them on ImageNet at a resolution of 128x128 pixels. The results are impressive: their BigGANs achieve an Inception Score (IS) of 166.3 and a Frechet Inception Distance (FID) of 9.6. These scores represent significant improvements over the previous best IS of 52.52 and FID of 18.65. Overall, this study demonstrates that by training GANs at a larger scale and incorporating orthogonal regularization and the truncation trick, it is possible to generate high-fidelity natural images with increased diversity from complex datasets like ImageNet. These findings contribute to advancements in generative image modeling and have implications for various applications in computer vision and artificial intelligence.