We combine Grad-CAM with existing fine-grained visualizations to create a high-resolution class-discriminative visualization and apply it to image classification, image captioning, and visual question answering (VQA) models, including ResN…