Stylizing Images Using Generative Adversarial Networks

This paper, addresses the very real challenge of changing photographic images of reality into corresponding cartoon images by working with advanced GANs. Working on the foundation established for CartoonGAN, it was trained over an extensive dataset that exceeded more than 45,000 images derived from the works of Makoto Shinkai at Studio Ghibli, regarded as among the greatest in his sector. Apart from CartoonGAN, StyleGAN3 and CycleGAN are considered as benchmarks for results. These models are assessed based on the retention of content, sharpness of the edges, viewership engagement, different variations in the output, quality of art, and uniformity of style. Qualitative survey is conducted for 150 participants and asked to rank the images generated by the model for all the metrics mentioned. The model performs well within the scope of the purposes without the necessity of an extensive exhaustive search, as this qualitative survey distinctly gauges the success of the proposed model against all the existing techniques, indicating the ability of the model and the remnant areas that require improvement. The results demonstrate that the approach possesses better content retention and style consistency than other methods since it produces plausible and high-quality cartoon-to-photo transformations. For future work, the model will be further fine-tuned and trained on the increased dataset’s size for enhancement of robustness and generalizing the performance altogether.