We propose a method to extrapolate the 360◦ field of view of a single input image. To do so, we first propose improvements to adapt an existing GAN-based in-painting architecture for out-painting on panoramic image representations.

Our method obtains state-of-the-art results and outperforms previous methods on standard image quality metrics. One major limitation of this out-painting method, however, is that it affords limited control over the synthesized content beyond sampling a different style. To circumvent this issue, we propose a novel guided co-modulation framework, which drives the image generation process with a common pre-trained discriminative model. Doing so maintains the high visual quality of the generated panoramas while enabling a user to control the semantic content of the extrapolated field of view. We demonstrate the state-of-the-art results of our method on field of view extrapolation both qualitatively and quantitatively and provide a thorough analysis of our novel editing capabilities. Finally, we demonstrate that this benefits the photorealistic virtual insertion of highly glossy objects in photographs.