Large Scale GAN Training for High Fidelity Natural Image Synthesis

This paper scales up Generative Adversarial Network training to the largest extent attempted at the time in order to generate high-resolution, diverse images from complex datasets such as ImageNet. The authors study the training instabilities that arise specifically at this scale, and find that applying orthogonal regularization to the generator makes it amenable to a 'truncation trick.' This trick reduces the variance of the generator's input, giving fine control over the trade-off between sample fidelity and variety.

The resulting models, called BigGANs, set a new state of the art in class-conditional image synthesis. Trained on ImageNet at 128x128 resolution, they achieve an Inception Score of 166.5 and a Frechet Inception Distance of 7.4, substantially improving over the previous best of 52.52 IS and 18.6 FID. This showed that scale, combined with targeted regularization, could dramatically raise the quality of generated natural images.