Where is the SceneVTG-benchmark?

Hi! I was wondering about the 'SceneVTG-benchmark' mentioned in the paper—where can I find it? I noticed that the SCUT-EnsText dataset only contains erased images, and I couldn’t locate the corresponding prompts. Also, after a quick look at SCUT-EnsText, I found 813 items, including English, Chinese, and even face erasures. Could you clarify how the 646 image pairs mentioned in the paper were selected from these? Thanks so much!