SalArt-VQA

SalArt-VQA SalArt-VQA is a closed-set visual question answering benchmark for salient artifact understanding in AI-generated images. Dataset Structure Split: test Rows: 950 Artifact images: 475 Clean real-image references: 356 Paired generated artifact-free counterparts: 119 Each row is self-contained. image path and q3 overlay image path are relative paths inside this dataset repository. image and q3 overlay image mirror those paths for Hugging Face image… See the full description on the dataset page:

Source: Hugging Face Hub (salartvqa/SalArt-VQA). Metadata imported from the dataset’s Hub tags.