Generative Adversarial Network Based Advanced Framework for High Resolution Multi-Style Cartoon Image Synthesis in Animation and Game Development
M. Kathiravan, A. Ramesh, R. Buvanesvari, S. Sreesubha, A. Irumporai, Raju D. Venkataramana
Abstract
Industries like entertainment, education, and advertisements require cartoon image generation as a most powerful business tool now a days. Classic systems such as texture synthesis and edge detection were largely unable to generate more natural images with an arbitrary artistic style. This research overcomes these limitations by presenting a Generative Adversarial Network (GAN)-based framework for high-resolution, style transferable cartoon picture generation. The use of GAN stems from the generator's creation of images and the discriminator's perception of which images it created, both of which aim to minimize and maximize costs, thereby increasing the diversity of outputs over time. The model was first trained with 40,000 images that were classified according to the anime-GAN dataset and then tested on 10,000 different anime images. Proposed a novel loss function that consists of both adversarial and content loss, allowing the generated images to preserve their structural relationships while providing better quality and style variety. Evaluation on Fréchet Inception Distance (FID) revealed that our model outperforms existing methods with a large margin both in image quality and adaptive diversity. It is fast, efficient, and versatile; it can generate cartoon images in a wide range of styles with no need for retraining. The proposed approach has numerous implications in animation and game development as a tool to generate cartoon images of high diversity and quality quickly, alleviating some limitations of existing traditional methods and past machine learning models.