The effect of diffusion model on image generation and performance analysis

This thesis investigates the impact and performance of the diffusion model on image generation, comparing it with Generative Adversarial Networks (GANs) and other diffusion models. The aim of the research is to evaluate the performance of models used in image generation, focusing on energy efficiency, processing times, and image quality. The study emphasizes the SinGAN, SinDDM, and SinFusion models and examines their performance at different depths. The study utilizes 21 images selected from the Places50 dataset, which are evaluated across the three models. The performance of the models is measured using metrics such as Sliced Frechet Inception Distance (SIFID), processing time, energy consumption, and memory usage. The findings indicate that the SinGAN model stands out with low processing times and energy consumption, although it shows some disadvantages due to its high parameter count. The SinDDM model, while capable of producing high-quality images, presents challenges in practical applications due to its very high processing times and energy consumption. The xiv SinFusion models demonstrate balanced performance at different depths, achieving a compromise between energy efficiency and processing times. In conclusion, the study provides significant insights into the performance of diffusion models and GANs, emphasizing the importance of balancing energy efficiency and image quality. Future research should test the performance of models on more complex and diverse datasets and explore optimization techniques to enhance energy efficiency. The findings underscore the importance of selecting the appropriate model based on application requirements.