Pixel-wise Conditioned Generative Adversarial Networks for Image Synthesis and Completion

Generative Adversarial Networks (GANs) have proven successful for\nunsupervised image generation. Several works have extended GANs to image\ninpainting by conditioning the generation with parts of the image to be\nreconstructed. Despite their success, these methods have limitations in\nsettings where only a small subset of the image pixels is known beforehand. In\nthis paper we investigate the effectiveness of conditioning GANs when very few\npixel values are provided. We propose a modelling framework which results in\nadding an explicit cost term to the GAN objective function to enforce\npixel-wise conditioning. We investigate the influence of this regularization\nterm on the quality of the generated images and the fulfillment of the given\npixel constraints. Using the recent PacGAN technique, we ensure that we keep\ndiversity in the generated samples. Conducted experiments on FashionMNIST show\nthat the regularization term effectively controls the trade-off between quality\nof the generated images and the conditioning. Experimental evaluation on the\nCIFAR-10 and CelebA datasets evidences that our method achieves accurate\nresults both visually and quantitatively in term of Fr\\'echet Inception\nDistance, while still enforcing the pixel conditioning. We also evaluate our\nmethod on a texture image generation task using fully-convolutional networks.\nAs a final contribution, we apply the method to a classical geological\nsimulation application.\n