The aim of this project was to develop a Generative Adverserial Network that could generate completely new faces from random noise by completely learning the underlying probability distribution of CelebA dataset.The CelebA dataset contains over 200,000 celebrity images with annotations.The system can learn and separate different aspects of an image unsupervised; and enables intuitive, scale-specific control of the synthesis.The network architecture followed the standard DCGAN DCGAN paper architecture with slight modifications done.As you can see,the model performs quite well with the generated faces well formed and distinguishable.

*The generated samples are largely white faces due to tha lack in diversity of the images belonging to different ethnicities.
*The model was quite small with a dept of 512,thus detailed features of eyes,ears etc couldnt be modelled acccurately.
*The number of epochs,the dept of the layers should be increased to achieve better results.