Exploring Generative Adversarial Networks for Text-to-Image Generation with Evolution Strategies

Text-to-image generation has achieved impressive results, featuring a variety of models trained on extensive datasets comprising text-image pairs. However, some methods depend on pre-trained models, using gradient-based approaches to update latent vectors in the latent space. In this work, we propose the use of Covariance Matrix Adaptation Evolution Strategy to explore the latent space of a Generative Adversarial Network. Our experimental study compares our approach with gradient-based and hybrid strategies, using diverse text inputs for image generation. We adapt an evaluation method that projects the generated samples into a two-dimensional grid to assess the diversity of the distribution. Results evidence that the evolutionary method produces more diverse samples across different grid regions, while the hybrid method combines gradient-based and evolutionary approaches, enhancing result quality.