Three Towers: Flexible Contrastive Learning with Pretrained Image Models | Read Paper on Bytez