Text to fashion image synthesis Via CW-ControlGAN

Text-to-image generation is called the method that translates human textual descriptions in the form of keywords or sentences into images with the same semantic content as text. The generated image in this field should match the meaning and content of the text and also have an acceptable quality. In the last few years, the use of Generative adversarial networks has made significant progress in increasing the visual realism, diversity, and semantic harmony of the generated images with their corresponding texts. Nevertheless, this research field is still facing many challenges. The research done is an attempt to develop a new architecture to compete with the state-of-the-art models. The research about text-to-image generation is done using two approaches: defining a loss function and then changing architecture. In the first approach, to extract more detailed features from the desired text and the generated image by the model, multi-mode transformer models that are trained on both the image and the text are used, and specifically, in this research, the text and image encoders of the CLIP model are used. The second approach deals with adding a wavelet loss function to the objective function to better train and improve the performance of the generator network in generating generated images. The obtained results show that a combination of two approaches produces good-quality images and better-attended text descriptions. The datasets used in this research are FashionGen and CUB Birds.