UniJEPA Unifies Image and Video World Modeling in One Latent Space, Planning Tens of Times Faster Than Generative World Models

arXiv 2608.07409·medium signalXBlueskyLinkedInCopy link

JEPA research has fragmented into incompatible recipes — I-JEPA predicts masked image parts, Image World Models predict photometric transformations, and V-JEPA 2 / DINO-World / DINO-WM predict future temporal states — each with separate encoders, predictors, and anti-collapse regularizers. UniJEPA learns photometric and temporal prediction jointly in one shared latent space under a single end-to-end objective (next-embedding loss plus a Gaussian regularizer) that is provably anti-collapse and trains from raw pixels with no EMA, stop-gradient, or pretrained encoder. It matches or surpasses task-specific JEPAs with one loss hyperparameter and plans up to tens of times faster than generative world models at comparable accuracy.