Aesthetic Image Captioning with Saliency Enhanced MLLMs