it is actually demonstrated that The easy pre-education process of predicting which caption goes with which graphic is an successful and scalable way to know SOTA impression representations from scratch on the dataset https://k2spiceshop.com/product/liquid-k2-on-paper-online/