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

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