Sketch-Guided Scenery Image Outpainting

The outpainting results produced by existing approaches are often too random to meet users' requirement. In this work, we take the image one step forward allowing users harvest personal custom using sketches as guidance. To end, propose an encoder-decoder based network conduct sketch-guided outpainting, where two alignment modules adopted impose generated content be realistic and consistent with provided sketches. First, apply a holistic module make synthesized part similar real from global view. Second, reversely produce encourage them ground-truth ones sketch module. way, learned generator will imposed pay more attention fine details sensitive guiding our knowledge, work is first attempt explore challenging yet meaningful conditional scenery outpainting. We extensive experiments on collected benchmarks qualitatively quantitatively validate effectiveness of approach compared other state-of-the-art generative models.