TextControlGAN: Text-to-Image Synthesis with Controllable Generative Adversarial Networks

Abstract

Generative adversarial networks (GANs) have demonstrated remarkable potential in the realm of text-to-image synthesis. Nevertheless, conventional GANs employing conditional latent space interpolation and manifold interpolation (GAN-CLS-INT) encounter challenges in generating images that accurately reflect the given text descriptions. To overcome these limitations, we introduce TextControlGAN, a controllable GAN-based model specifically designed for text-to-image synthesis tasks. In contrast to traditional GANs, TextControlGAN incorporates a neural network structure, known as a regressor, to effectively learn features from conditional texts. To further enhance the learning performance of the regressor, data augmentation techniques are employed. As a result, the generator within TextControlGAN can learn conditional texts more effectively, leading to the production of images that more closely adhere to the textual conditions. Furthermore, by concentrating the discriminator’s training efforts on GAN training exclusively, the overall quality of the generated images is significantly improved. Evaluations conducted on the Caltech-UCSD Birds-200 (CUB) dataset demonstrate that TextControlGAN surpasses the performance of the cGAN-based GAN-INT-CLS model, achieving a 17.6% improvement in Inception Score (IS) and a 36.6% reduction in Fréchet Inception Distance (FID). In supplementary experiments utilizing 128 × 128 resolution images, TextControlGAN exhibits a remarkable ability to manipulate minor features of the generated bird images according to the given text descriptions. These findings highlight the potential of TextControlGAN as a powerful tool for generating high-quality, text-conditioned images, paving the way for future advancements in the field of text-to-image synthesis.