One of the key factors that are responsible for the success of deep learning is the methods or group of methods. In the recent past, Autoencoder (AE) and Generative Adversarial Network (GANs) has grown vast popularity in deep learning community. Autoencoder (AE) and GAN is employed to generate images in various domains like as computer vision, semantic segmentation and medical field. In this paper we compare and implement the three autoencoders model simple autoencoder, vanilla autoencoder and convolutional autoencoder with different architecture. The first autoencoder, simple autoencoder with only bottleneck layer, second autoencoder is with one hidden layer and bottleneck layer and third autoencoder is the convolutional autoencoder. We use convolutional layers in convolutional autoencoder and in generative adversarial network, that is better to capture the spatial information in image rather than using one or more hidden layer as in simple autoencoder and vanilla autoencoder. We have generated the images using these autoencoders and compare the training loss, validation loss and accuracy.Image is also generated via Fully Connected Generative Adversarial Network (FCGAN) for a particular batch size.