How can adversarial training be used to reduce unrealistic image generation in deep learning-based image models

Question

Can I get a good explanation of how adversarial training can be used to reduce unrealistic image generation in deep learning-based image models?