[Paper Review] Progressive Growing of GANs for Improved Quality, Stability, and Variation

Tero Karras, Timo Aila+2 more|arXiv (Cornell University)|Oct 27, 2017

Generative Adversarial Networks and Image SynthesisComputer Science2,871 citations

TL;DR

This paper introduces progressive growing of both generator and discriminator to stabilize GAN training and produce high-resolution images, with demonstrated gains in quality and variation, including CelebA-HQ 1024^2 and CIFAR-10 unsupervised score of 8.80.

ABSTRACT

We describe a new training methodology for generative adversarial networks. The key idea is to grow both the generator and discriminator progressively: starting from a low resolution, we add new layers that model increasingly fine details as training progresses. This both speeds the training up and greatly stabilizes it, allowing us to produce images of unprecedented quality, e.g., CelebA images at 1024^2. We also propose a simple way to increase the variation in generated images, and achieve a record inception score of 8.80 in unsupervised CIFAR10. Additionally, we describe several implementation details that are important for discouraging unhealthy competition between the generator and discriminator. Finally, we suggest a new metric for evaluating GAN results, both in terms of image quality and variation. As an additional contribution, we construct a higher-quality version of the CelebA dataset.

Motivation & Objective

• Motivate and address instability and limited variation in GAN training at high resolutions.
• Propose a training regime that progressively grows both generator and discriminator from low to high resolution.
• Introduce simple techniques to increase output variation without extra hyperparameters.
• Showcase high-quality megapixel generation (CelebA at 1024^2) and strong CIFAR-10 unsupervised inception scores.
• Offer practical implementation details to discourage unhealthy competition between G and D and propose a new evaluation metric for quality and variation.

Proposed method

• Progressively grow G and D from low (4x4) to high (1024x1024) resolutions by adding layers and fading them in during transitions.
• Use equalized learning rate to keep per-layer learning speeds balanced across layers.
• Apply pixelwise feature vector normalization in the generator to constrain signal magnitudes.
• Incorporate a minibatch stddev feature in the discriminator to promote variation without extra learned parameters.
• Employ a simplified minibatch discrimination concept that does not add learnable parameters or hyperparameters.
• Utilize a multi-scale statistical similarity metric (SWD) based on a Laplacian pyramid to assess image distribution similarity to the training set.

Experimental results

Research questions

• RQ1 Can progressive growth of both generator and discriminator stabilize GAN training at high resolutions?
• RQ2 Does progressive growing improve image quality and training speed compared to non-progressive GANs?
• RQ3 How can variation in generated images be increased without introducing additional hyperparameters or complex regularizers?
• RQ4 What metrics best quantify both quality and diversity of GAN outputs across scales?

Key findings

• Progressive growing significantly speeds up training and stabilizes high-resolution image generation (e.g., up to 1024^2).
• Achieves improved convergence and higher-quality images across CelebA, LSUN, and CIFAR-10 settings.
• Minibatch stddev and equalized learning rate contribute to reduced mode collapse and more varied outputs.
• A higher-quality CelebA dataset (CelebA-HQ) at 1024^2 is created to support high-res evaluation.
• CIFAR-10 unsupervised inception score reaches 8.80, surpassing prior unsupervised results.
• SWD-based multi-scale evaluation aligns with qualitative improvements and distinguishes configurations beyond MS-SSIM.