What’s that Pokémon: Image Synthesis Using a DCGAN Architecture

In recent years, deep learning algorithms have evolved considerably, displaying exceptional performance in fields such as computer vision, image synthesis, natural language processing, and more. One of the most well-studied such algorithms is GAN (Generative Adversarial Network), most commonly used in image synthesis. In this paper, we explore a variant of GAN known as DCGAN - deep convolutional GAN, and apply it to a Pokémon image dataset, to generate plausible-looking new Pokémon. The dataset is first preprocessed into 128x128 images and augmented with additional samples derived from the original data for increased variance. Afterward, a DCGAN with several beneficial improvements (weight initialization, label smoothing, instance noise) and recommended parameters is trained on the dataset. The model’s number of layers, layer hyperparameters, and loss functions are tuned until the synthesized images display a satisfactory resemblance to the original samples. Lastly, several super-resolution algorithms are tested on the generated outputs to upscale them to a 512 × 512 resolution.