This is an implementation of Co-Modulated Generative Adversarial Networks (CoModGAN), a novel GAN architecture for image completion that co-modulates both conditional and stochastic style representations. The model bridges image-conditional and modulated unconditional generative architectures to handle large-scale missing regions that prior methods struggle with. It also introduces Paired/Unpaired Inception Discriminative Scores (P-IDS/U-IDS) as quantitative metrics for measuring perceptual fidelity of inpainted images.