LocoGAN — Locally convolutional GAN

摘要

We propose LocoGAN — a fully convolutional GAN model with latent space given by image-like noises of possibly different resolutions. Its learning procedure is local and processes not the whole image-like noises but only the sub-images of a fixed size. Consequently, LocoGAN produces images of arbitrary dimensions, which we present using the LSUN bedroom data set. By leveraging local learning and incorporating the position channels, our model gains an uncommon ability to generate fully periodic (e.g., cylindrical panoramic images) or almost periodic ”infinitely long” images.