LeCun clarifies: DINO is a JEPA variant and most JEPA papers are just fine-tunes

ylecun · x · 2026-11-08

Yann LeCun posted a detailed technical clarification on the JEPA debate:

Image/video SSL methods differ by whether they use reconstruction; joint embedding architectures (JEA vs JEPA) may or may not have a predictor, and also differ in anti-collapse methods and data augmentation.
DINO is technically a kind of JEPA (a JEA); its distinctives are the anti-collapse criterion (contrastive clustering + stop-gradient/EMA) and task-specific image-transformation augmentations.
V-JEPA 2.1, like DINO, comes from Meta-FAIR, built with large engineering teams, proprietary datasets, considerable compute, and extensive benchmark tuning.
Many papers with "JEPA" in the title (LeJEPA, LeWM, etc.) are applications or fine-tuned versions of DINO, V-JEPA, or I-JEPA rather than fundamentally new work.

More from Models