3D Face Image Inpainting with Generative Adversarial Nets

Figure 1

Our training starts with both the generator G and discriminator D having a low spatial resolution of m × n pixels. As the training advances, we incrementally add layers to G and D thus increasing the spatial resolution of the generated images. All existing layers remain trainable throughout the process. Here, m × n refers to convolutional layers operating on M × N spatial resolution [16]. This allows stable synthesis in high resolution and speeds up the training considerably.