Effective methods for detecting and monitoring underwater trash are necessary since the growth of it poses a serious hazard to marine ecosystems. The efficient application of computer vision for the visual detection of marine debris is hindered by the data scarcity problem in underwater debris datasets. Generative Adversarial Networks (GANs) have shown themselves to be an effective tool in generating realistic pictures, providing a potential solution to scarce dataset augmentation. However, due to data insufficiency and low quality of the underwater debris images, performance of GANs often do not provide an improvement of the final model learning, thus investigation on dataset design and augmentation is of the utmost importance. In this work a specific GAN variation called Deep Convolutional GAN (DCGAN) is employed for the purpose of generating synthetic dataset of undersea trash, and its performance is investigated on three differently designed datasets, including two publicly available datasets from both, underwater environment and terrestrial environment, and one dataset created through text-to-image generator. The experimental results provide novel insights and recommendation for further research. The paper provides detailed information on dataset methodology, training techniques, and the architecture of the model developed using Python.