Unifying Conditional and Unconditional Semantic Image Synthesis with OCO-GAN

Generative image models have been extensively studied in recent years. In the unconditional setting, they model marginal distribution from unlabelled images. To allow for more control, synthesis can be conditioned on semantic segmentation maps that instruct generator position of objects image. While these two tasks are intimately related, generally isolation. We propose OCO-GAN, Optionally COnditioned GAN, which addresses both a unified manner, with shared network either or directly latents. Trained adversarially an end-to-end approach discriminator, we able to leverage synergy between tasks. experiment Cityscapes, COCO-Stuff, ADE20K datasets limited data, semi-supervised and full data regime obtain excellent performance, improving over existing hybrid generate without conditioning all settings. Moreover, our results competitive better than state-of-the art specialised conditional models.