
Figure 1: Our C-JEPA achieves faster and betterconvergencethanI-JEPA. Unsupervised learning ofvisual representations has recently seen remarkable progress, primarily due to the development of innovative architectures and strategies that exploit unlabeled imagery.