USIS: Unsupervised Semantic Image Synthesis

Semantic Image Synthesis (SIS) is a subclass of I2I (I2) translation where photorealistic image synthesized from segmentation mask. SIS has mainly been addressed as supervised problem. However, state-of-the-art methods depend on massive amount labeled data and cannot be applied in an unpaired setting. On the other hand, generic frameworks underperform comparison. In this work, we propose new framework, Unsupervised (USIS), first step towards closing performance gap between paired settings. We design simple effective learning scheme that combines fragmented benefits cycle losses relationship preservation constraints. Then, make discovery that, contrary to translation, discriminator crucial for label-to-image translation. To end, with wavelet-based encoder decoder reconstruct real images. The self-supervised reconstruction loss prevents overfitting few wavelet coefficients. test our methodology 3 challenging datasets set standard SIS. generated images demonstrate significantly better diversity, quality multimodality.