it can be shown that the simple pre-coaching process of predicting which caption goes with which impression is definitely an economical and scalable way to learn SOTA picture representations from scratch on the dataset https://k2spiceshop.com/product/liquid-k2-on-paper-online/