# Progressive Growing of GANs for Improved Quality, Stability, and Variation

**Authors:** Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen  
**arXiv:** [1710.10196](https://arxiv.org/abs/1710.10196)  
**Date:** 2017-10-27  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1710.10196v3/

## Key points

- Progressive growing of GANs: a novel training methodology
- Growing both generator and discriminator networks during training
- Adding new layers to model increasingly fine details
- Speeding up and stabilizing the training process
- Generating high-quality images at impressive resolutions (e.g., 1024^2)
- Increasing variation in generated images
- Achieving record-breaking inception score in unsupervised CIFAR10 dataset experiments (8.80)
- Implementation details to discourage unhealthy competition between networks
- New metric for evaluating GAN results based on image quality and variation
- Construction of a higher-quality version of the CelebA dataset for future research
- Improved image quality, stability, and variation compared with existing methods

## Summary

In their paper titled "Progressive Growing of GANs for Improved Quality, Stability, and Variation," authors Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen introduce a novel training methodology for generative adversarial networks (GANs). The main idea behind this approach is to progressively grow both the generator and discriminator networks during training. Starting from a low resolution, new layers are added to model increasingly fine details as the training progresses. This progressive growing technique not only speeds up the training process but also greatly stabilizes it. As a result, the authors are able to generate high-quality images that were previously unattainable. For example, they successfully produce CelebA images at an impressive resolution of 1024^2. Furthermore, the authors propose a simple yet effective method to increase the variation in generated images. They achieve a record-breaking inception score of 8.80 in unsupervised CIFAR10 dataset experiments. The paper also delves into several crucial implementation details that help discourage unhealthy competition between the generator and discriminator networks. These details contribute to maintaining stability and improving overall performance. Additionally, the authors suggest a new metric for evaluating GAN results based on image quality and variation. This metric provides a comprehensive assessment of GAN outputs. As an additional contribution to the field, the authors construct a higher-quality version of the CelebA dataset which serves as valuable input for future research and development in generative models. Overall, this paper presents an innovative approach to training GANs that leads to improved image quality, stability, and variation compared with existing methods. The proposed methodology offers valuable insights into addressing challenges associated with GAN training while achieving remarkable results in terms of image generation capabilities.