Facial Image Generation with Limited Training Data

Deep learning models have a wide number of applications including generating realistic-looking images. These typically require lots data, but we wanted to explore how much quality is sacrificed by using smaller amounts data. We built several and trained them at different dataset sizes, then assessed the generated images with widely used FID measure. As expected, measured an inverse correlation -0.7 between image training set size. However, observed that small-training-set results had problems not detectable this experiment. therefore present experimental design for follow-up study would further lower limits experiments are important bringing us closer understanding data needed train successful generative model.