Дослідження генеративно-змагальних моделей для синтезу тексту в зображення

Abstract

This work is devoted to the study of the architecture and work of GAN networks for text-to-image synthesis. The paper highlights some common problems that a developer may encounter when using vanilla GAN. Ways to solve these problems in the form of GAN models for generating high quality images have been presented. We critically examine current strategies for evaluating text-to-image synthesis models, identify the drawbacks and merits of each model. As a result, the best solution in the form of using a suitable model for high-quality image generation from text descriptions is selected.