Stabilizing GANs with Soft Octave Convolutions
This addresses stabilization issues in GANs for image generation, offering an incremental improvement that is orthogonal to existing methods.
The paper tackles the problem of stabilizing GAN training and reducing mode collapse by proposing a novel convolution scheme that splits filters into high and low frequency parts, shifting weight updates during training to learn coarse structures first, which reduces frequency artifacts in generated images.
Motivated by recently published methods using frequency decompositions of convolutions (e.g. Octave Convolutions), we propose a novel convolution scheme to stabilize the training and reduce the likelihood of a mode collapse. The basic idea of our approach is to split convolutional filters into additive high and low frequency parts, while shifting weight updates from low to high during the training. Intuitively, this method forces GANs to learn low frequency coarse image structures before descending into fine (high frequency) details. We also show, that the use of the proposed soft octave convolutions reduces common artifacts in the frequency domain of generated images. Our approach is orthogonal and complementary to existing stabilization methods and can simply be plugged into any CNN based GAN architecture. Experiments on the CelebA dataset show the effectiveness of the proposed method.