Geometric Decoupling: Diagnosing the Structural Instability of Latent

Latent Diffusion Models (LDMs) achieve high-fidelity synthesis but suffer from latent space brittleness, causing discontinuous semantic jumps during editing. We introduce a Riemannian framework to diagnose this instability by analyzing the generative Jacobian, decomposing geometry into $\mathit{\text{Local}}\mathit{\text{Scaling}}$ (capacity) and $\mathit{\text{Local}}\mathit{\text{Complexity}}$ (curvature). Our study uncovers a $\mathbf{\text{“Geometric}}\mathbf{\text{Decoupling"}}$: while curvature in normal generation functionally encodes image detail, OOD generation exhibits a functional decoupling where extreme curvature is wasted on unstable semantic boundaries rather than perceptible details. This geometric misallocation identifies ``Geometric Hotspots" as the structural root of instability, providing a robust intrinsic metric for diagnosing generative reliability.