New method trains vision-language models without pairing images

Original: RT @lukaskuhn77: 🔥 We introduce LeVLJEPA: the first fully non-contrastive end-to-end vision-language pretraining method competitive with CL…

Source: x.com ↗