ARGAN: Fast Converging GAN for Animation Style Transfer

Abstract

Transformation of real images to the animated image is one of the most challenging tasks in artistic style transfer. In this paper, using a novel architecture for Generative Adversarial Networks (GANs), a faster and more accurate result for style transfer is achieved. There are three common problems regarding animation style transfer. First, the original content of an image is lost during the generation of new images by the network. Second, the generated image does not have an apparent animated style. Finally, the networks are not fast enough, and they require a large amount of memory to process the images. In this paper, ARGAN, a lightweight and fast GAN network for animation style transfer, is proposed. To enhance the quality of the output images, three loss functions related to grayscale style, content style, and reconstruction of the color spectrum in each image are proposed. Also, the training phase of this method does not require paired data. The proposed method transforms real-world images into animated style images significantly faster than similar methods.