TEXT DESCRIPTION TO IMAGE GENERATION USING GENERATIVE ADVERSARIAL NETWORK

Abstract

The progression of translating text into images has been an imperative topic of research. The significant challenges arise from translating visual to textual information and vice versa. High-quality images can be generated from text using a Generative Adversarial Network (GAN), however, there are challenges associated with accurately portraying the content of the sentence provided to the model. Text-to-image conversion strategies can produce examples that closely reflect the descriptions' intended meaning. The user descriptions may however lack crucial details. To create the conditioned text descriptions, this study employs an Attention-Generative Adversarial Network to generate 256*256-pixel images that are image-sensitive. In the initial phase of GAN sketches, the input text descriptions serve solely to inform the basic form and color scheme of the devices. The information gleaned from the first stage, along with the textual descriptions, is fed into a GAN which generates images with high resolution and realistic detail. The conditional GAN training may be stabilized using conditioning augmentation, and the generated samples can have higher quality. Using Style based Generator, samples for each style of the image can be drawn. The proposed system can generate photorealistic visuals of an object when the user inputs the textual descriptions in the application’s GUI.

Keywords

• generative adversarial network (gan)
• attention generative adversarial network
• generator
• discriminator
• style based generator
• text to image.