JEPA models predict latent representations rather than pixels, but a standard predictor returns one answer even when an image admits several plausible completions. This project asks what changes when that prediction becomes a distribution: it compares mixture, autoregressive, flow-matching, and diffusion heads on the same visual JEPA backbone, probing uncertainty without giving up semantic representation learning.