Abstract
In this work, we develop a method to generate infinite high-resolution imageswith diverse and complex content. It is based on a perfectly equivariantgenerator with synchronous interpolations in the image and latent spaces.Latent codes, when sampled, are positioned on the coordinate grid, and eachpixel is computed from an interpolation of the nearby style codes. We modifythe AdaIN mechanism to work in such a setup and train the generator in anadversarial setting to produce images positioned between any two latentvectors. At test time, this allows for generating complex and diverse infiniteimages and connecting any two unrelated scenes into a single arbitrarily largepanorama. Apart from that, we introduce LHQ: a new dataset of \lhqsizehigh-resolution nature landscapes. We test the approach on LHQ, LSUN Tower andLSUN Bridge and outperform the baselines by at least 4 times in terms ofquality and diversity of the produced infinite images. The project page islocated at https://universome.github.io/alis.