A Comparative Analysis of DCGAN and WGAN in Logo Generation

This research paper explores the impact of Generative Adversarial Networks (GANs) on logo generation, focusing on three variants: Deep Convolutional GAN (DCGAN), Wasserstein GAN (WGAN), and Conditional GAN (cGAN). The study evaluates their strengths and weaknesses, aiming to contribute insights to the ongoing discourse on GANs. Results indicate that DCGAN excels in capturing fine details but faces mode collapse, limiting diversity, while WGAN demonstrates enhanced training stability, consistently delivering diverse and visually appealing logos. Numerical results reveal a mean rating of 3.173/5 across sample sets, with approximately 14/64 images deemed useful, resulting in a usability rating of 2.2/10. The survey findings, based on assessments from 57 participants, consistently rate DCGAN-generated images slightly higher in terms of visual appeal and usage preferences compared to those produced by WGAN. While DCGAN exhibits superior image quality, WGAN offers more variety. This suggests a trade-off between variety and image quality, underscoring the specific strengths of each model in meeting subjective criteria for visual excellence.