High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs

This work presents a method for turning semantic label maps into high-resolution, photo-realistic images using conditional generative adversarial networks. Previous conditional GANs produced low-resolution results still far from realistic, so the authors introduce a novel adversarial loss together with new multi-scale generator and discriminator architectures, enabling visually appealing results at 2048x1024 resolution.

The framework extends beyond synthesis to interactive visual manipulation. By incorporating object instance segmentation, it supports operations like removing, adding, or changing the category of objects, and a further method produces diverse results from the same input so users can interactively edit object appearance. Human opinion studies show the approach significantly outperforms existing methods, advancing both the quality and resolution of deep image synthesis and editing.