MITIGATING MODEL COLLAPSE IN GENERATIVE MODELS THROUGH HYBRID APPROACHES: A FOCUS ON VAE-GAN INTEGRATION

Abstract

Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) are examples of generative models that have advanced the field of artificial intelligence by producing realistic synthetic data. However, they often suffer from model collapse, which dramatically reduces the diversity and quality of generated samples. This study explores the possibility of integrating GANs with VAEs (VAEGAN) to mitigate model collapse, identifying key success factors such as architecture, training objectives, and hyper-parameters. The study demonstrates that the VAEGAN approach improves output diversity and quality, achieving an FID score of 35 and an IS score of 4.8, outperforming GAN (FID: 50, IS: 4.2) and VAE (FID: 20, IS: 3.5). The results show that VAEGAN achieves a loss value of 0.6, outperforming GAN's loss value of 0.8, while VAE shows the lowest loss value of 0.4. The study evaluates the performance of VAEGAN using the MNIST dataset, showcasing its ability to generate high-quality images with improved diversity. The findings suggest that the integration of VAE and GAN can effectively mitigate model collapse, leading to more robust and reliable generative models. The study confirms the existence of model collapse and highlights the benefits of VAE-GAN integration in strengthening generative modeling, paving the way for further research in this area. With the promising results achieved in this study, the VAEGAN approach has potential applications in various fields, including image and video generation, data augmentation, and anomaly detection. This study contributes to the development of more advanced generative models, enabling the generation of high-quality and diverse samples with potential applications in various fields.