Abstract
In this study, we replicate the text-adaptive generative adversarial network proposed by the paper "Text-Adaptive Generative Adversarial Networks: Manipulating Images with Natural Language", NeurIPS 2018. We do this for two reasons: first, image modification with natural language is an interesting and challenging problem, and second, replicating studies can provide additional information about the validity and generalizability of the original authors' claims. Our replication followed the specifications of the network given in the published study and supplementary materials (and not the publicly available code provided by the authors) and we trained our network using a dataset that contains 11,788 bird images. In addition to this, we use natural language captions, where 10 individual sentences are annotated for each image. We evaluated our network using L2 error and informal qualitative measures and found that our results were significantly worse than the results from the original study, due to issues in accurately and comprehensively recreating the network.
