A versatile conditional Wasserstein GAN in Python, designed and trained for human portrait recolorization.
Colorizing grayscale human portraits is an important task well suited for generative modeling, allowing for the realistic generation of new image data. The deep biases associated with many conventional machine learning datasets often come into view, considering the importance of fairly generating image data for individuals from diverse backgrounds. This project implements a conditional Wasserstein Generative Adversarial Network (cWGAN) to realistically colorize human portraits, trained on the AISegment.cn human portrait dataset. This work demonstrates that, despite the underlying biases in the training data, the network can be implemented and designed so that it performs well on diverse demographics of race, age, and gender, generating convincing output images. An extensive analysis is performed on images colorized through this network from a variety of human portrait datasets.
More information regarding the model and its analysis can be found in the written-report linked here.