In the paper THINGS-data, a multimodal collection of large-scale datasets for investigating object representations in human brain and behavior, it is said that
"For MRI, there were 100 separate test images that were a representative subset of the 200. Finally, there were 100 unique catch images that were created using the generative adversarial neural network BigGAN."