Enhanced Pix2Pix GAN for Sketch Colorization: Unveiling Multi-scale Discriminator with U - Net Architecture

In the field of image-to-image translation, with a particular emphasis on the intricate task of adding color to black-and-white sketches, this study presents an innovative enhancement to the Pix2Pix model. Our contribution lies in incorporating a sophisticated multi-scale discriminator, strategically designed to capture not only the finer details of line drawings but also the broader global structures inherent in the sketches. This unique approach significantly enhances the overall colorization process, resulting in exceptional outcomes that surpass traditional methods. Through meticulous experimentation on the available datasets of Anime and Pokemon sketches and their colored pairs images, the proposed enhanced Pix2Pix model demonstrates great potential by generating images that closely resemble the ground truth. Notably, our model outperforms the baseline Pix2Pix model in both quantitative and qualitative measures. On the Anime dataset, the proposed model surpasses its counterparts with a minimum difference of a minimum improvement of 32.09% in FID, 8% in SSIM, and 23.3% in PSNR respectively.