Optimizing and interpreting the latent space of the conditional text-to-image GANs

Abstract

Text-to-image generation intends to automatically produce a photo-realistic image, conditioned on a textual description. To facilitate the real-world applications of text-to-image synthesis, we focus on studying the following three issues: (1) How to ensure that generated samples are believable, realistic or natural? (2) How to exploit the latent space of the generator to edit a synthesized image? (3) How to improve the explainability of a text-to-image generation framework? We introduce two new data sets for benchmarking, i.e., the Good & Bad, bird and face, data sets consisting of successful as well as unsuccessful generated samples. This data set can be used to effectively and efficiently acquire high-quality images by increasing the probability of generating Good latent codes with a separate, new classifier. Additionally, we present a novel algorithm which identifies semantically understandable directions in the latent space of a conditional text-to-image GAN architecture by performing independent component analysis on the pre-trained weight values of the generator. Furthermore, we develop a background-flattening loss (BFL), to improve the background appearance in the generated images. Subsequently, we introduce linear-interpolation analysis between pairs of text keywords. This is extended into a similar triangular ‘linguistic’ interpolation. The visual array of interpolation results gives users a deep look into what the text-to-image synthesis model has learned within the linguistic embeddings. Experimental results on the recent DiverGAN generator, pre-trained on three common benchmark data sets demonstrate that our classifier achieves a better than 98% accuracy in predicting Good/Bad classes for synthetic samples and our proposed approach is able to derive various interpretable semantic properties for the text-to-image GAN model.