We derive Riemannian metrics from pretrained EBMs to compute data-aware geodesics. Our approach outperforms standard methods across datasets, offering a scalable solution for learning data geometry in high-dimensional spaces.