Abstract
Internal learning for single-image generation is a framework, where agenerator is trained to produce novel images based on a single image. Sincethese models are trained on a single image, they are limited in their scale andapplication. To overcome these issues, we propose a meta-learning approach thatenables training over a collection of images, in order to model the internalstatistics of the sample image more effectively. In the presented meta-learningapproach, a single-image GAN model is generated given an input image, via aconvolutional feedforward hypernetwork $f$. This network is trained over adataset of images, allowing for feature sharing among different models, and forinterpolation in the space of generative models. The generated single-imagemodel contains a hierarchy of multiple generators and discriminators. It istherefore required to train the meta-learner in an adversarial manner, whichrequires careful design choices that we justify by a theoretical analysis. Ourresults show that the models obtained are as suitable as single-image GANs formany common image applications, significantly reduce the training time perimage without loss in performance, and introduce novel capabilities, such asinterpolation and feedforward modeling of novel images.