Toward Robust Hyper-Detailed Image Captioning