Generation of Controlled Face Images using Curated Datasets and StyleGAN2

GANs have been researched and developed a lot in the past few years to advance the field of Image Generation and StyleGAN especially has been worked upon extensively to produce 3 successive versions. Despite this, due to its algorithmic design of combining or mixing of “styles”, there has so far been no way of definitively controlling the output of the images/faces produced by a trained StyleGAN model. In this paper, the proposed approach is, to control the sizes of facial features in the images generated by StyleGAN2 by feeding curated image data (produced via the discussed data pre-processing method) to the GAN model. The said data preprocessing approach comprises facial-feature size-ratio calculation, data discretization through binning and a unique querying approach. Testing has been performed to check whether the approach improves the quality of output images produced either in terms of metrics like FID/KID or in terms of time-efficiency by comparing the curated based training output to its non-curated counterpart. The generated output is also subjected to a verification process to estimate the levels of control obtained over the sizes of the facial features of generated images. It was found that the proposed data preprocessing methodology helps manipulate facial feature sizes in at most 70% of the generated images (based on sample testing). It is to be noted that these values were achieved using relatively primitive methods of image discretization and also with the GAN system being trained on dataset sizes which were smaller than recommended.