3.1. Generating a Multi-modal Training Dataset We combine the abilities of two large-scale pretrained models that operate on different modalities—a large lan- guage model [7] and a text-to-image model [51]—to gen- erate a multi-modal training dataset containing...