/Image_Generation_using_GANs

Generate new Faces by training DCGAN on CelebA dataset. Pytorch Implementation

Primary LanguageJupyter NotebookMIT LicenseMIT

Face-Generation

Face Generator Project is a part of Udacity Deep Learning Nanodegree online course

Goal

Generate new faces using Generative Adversarial Networks (GANs).
The model is trained on the CelebFaces Attributes Dataset (CelebA): Image of Training Set

It generates new human faces that look like this:
Image of Generated Faces

Recommended next steps

  • Create a deeper model and use it to generate larger (say 128x128) images of faces.
  • Read existing literature to see if you can use padding and normalization techniques to generate higher-resolution images.
  • Use WGAN - Wasserstein loss with Gradient penalty to mitigate problems like mode collapse(where generated faces have features from just few faces in training dataset and not from entire distribution of Celeb A dataset) and stabilise training process.