VFJ-Bench is a benchmark dataset designed to evaluate how well Vision-Language Models (VLMs) can detect and describe changes in edited images. The dataset contains original images along with their edited variants (both object editing and object deletion) at 4 different intensity levels, paired with detailed captions.
