Surrealistic-like Image Generation with Vision-Language Models