Pegasus Image Generating GAN

  • Category: ML model
  • Client: Durham University
  • Project date: April 2021


A DCGAN implemented in PyTorch used to turn images of planes and horses in CIFAR-10 into pegasus images The DCGAN was then experimented on to see how different modifications effect results.

STAR Breakdown

  • Situation: In order to see the effect of hybridizing different modifications on a DCGAN, I wanted to see how realistic I could get a picture of pegasus made with AI.
  • Task: I was asked to develop a ML model for creating pegasus images but was not instructed specifically how nor how I should experiment on it to try and tweak results.
  • Action: I first explored the data in the CIFAR-10 dataset and isolated the images of horses, planes and birds. I decided to use planes instead of birds as a method of implementing wings due to many bird images not having spread wings nor being white in colour. I then created the layered convolutions in the DCGAN and fed the images into it. The resulting image batches were promising and so I began tweaking various metrics to try and improve the result.
  • Result: My written report on the experiments discusses the experiments made and their effects as well as contains the final generated images.