Boosting GAN Performance Through Dataset Augmentation with Denoising Diffusion Models

Image generation using generative models is a new and challenging task, especially when the available training dataset is limited in diversity and size. Generative AI models require a huge corpus of data for training effectively. Limited data can cause problems such as overfitting and poor generalization, making it difficult to generate images. To Address this challenge, this work explores an innovative approach to integrate Denoising Diffusion Probabilistic models (DDPMs) and Generative Adversarial Networks (GANs).The aim of this work is to significantly enhance the quality and diversity of generated images by initially augmenting the training dataset and then utilizing the potential of GAN s for realistic image generation.In the proposed hybrid model, the DDPM acts as the initial step, performing data augmentation by generating additional samples.then the augmented dataset is passed through GAN. To optimize the model's performance, a careful analysis of varying augmentation ratios and training epochs was conducted. The Experimentation concluded that as the dataset size increases the models performance improved drastically on FID metric.On the CIFAR-I0 Dataset an 82.96% improvement was observed demonstrating the positive effect of augmentation. The findings highlight the effectiveness of combining diffusion-based data augmentation with GAN s in addressing challenges like limited data diversity and availability.