As generative models have made image editing easier, Image Difference Captioning (IDC), which describes the differences between two images, has become increasingly important. While existing IDC models have been successful with 3D rendered images, they have struggled with real-world images due to a lack of training data and the difficulty of capturing subtle differences in complex images. To address these challenges, this paper proposes a simple yet effective framework that adapts existing image captioning models to the IDC task and augments the IDC dataset. Specifically, we develop the BLIP2IDC model, which adapts BLIP2 to the IDC task and demonstrates superior performance compared to existing approaches using two streams. Furthermore, we propose a novel Syned1 dataset, which enhances the performance of IDC models through synthetic data augmentation.
