Ashley-ht/MDID

Dataset Card for Dataset Name

This is a multidimensional, diverse, and impartial dataset for face-swapping image quality diagnosis. Our data collection is based on the widely-used FFHQ dataset, which provides high-resolution facial images with diverse poses and illumination conditions. Importantly, FFHQ has become a de facto standard in face-swapping research due to its well-documented consent, licensing, and privacy compliance, making it suitable for large-scale data release and redistribution. By adopting FFHQ, we ensure that MDID adheres to established ethical and legal practices in portrait data usage, while maintaining compatibility with prior face-swapping pipelines. From FFHQ, we sample 1,800 images, maintaining a balanced 1:1 male-to-female ratio within each group. To construct MDID, we generate face-swapped versions of these 1,800 source images using six representative face-swapping methods. We select three GAN-based methods (SimSwap, InfoSwap, E4S) and three Diffusion-based methods (DiffSwap, Face-Adapter, REFace), all sourced from official releases. All models are deployed using official open-source weights obtained directly from the authors' repositories, ensuring reproducibility and consistency with the original implementations.