Gaussian-JEPA: Joint-Embedding Predictive Learning applied to 3D Gaussian Splats

shumpeiMaxwell · x · 2026-08-19

A new paper, Gaussian-JEPA, brings Joint-Embedding Predictive Learning (JEPA) to 3D Gaussian Splatting: instead of reconstructing raw Gaussian attributes, it predicts hidden Gaussian-token features from visible context. Per the TL;DR, this improves robustness to resampling, partial views, and Gaussian completion tasks.

Related event: Gaussian-JEPA Brings JEPA to 3D Gaussian Splatting(2 posts)→

Original post →