SAFID: Synthetically Augmented First Impression Data

We investigate whether adding synthetic face variations of new identities within the training sets, to train first impression prediction models, improves validation performance. We initially sample 4373 real andsynthetic faces each, from FFHQ and synthetic images generated by StyleGAN2-ADA, respectively. We then augment these sets using synthetic variations generated using our method (paper will be linked after acceptance).The 17492 generated samples from both these sets are appended to our training datasets (details are provided in the paper).