Fisher information metric reconstruction from a log-partition function

methods / Theoretical / Analytical / Geometric analysis / Fisher information metric reconstruction from a log-partition function

Fisher information metric reconstruction from a log-partition function

Techniqueadvanced

Reconstructs an information-geometric (Fisher) metric over a generative model's latent/noise space by treating the model as an implicit exponential-family log-partition function, validated against exactly-solvable statistical-physics systems (Ising, TASEP) before being applied to a real trained generative model to detect abrupt metric discontinuities (phase transitions) rather than smooth curvature.

Used in (1 observation)

A real Stable Diffusion 1.5 latent space shows a fractal structure of Fisher-metric phase transitions, with diverging Lipschitz constants at phase boundariesmeasured

structure: Curvature profile of the representation manifold · models: Stable Diffusion 1.5 (Dreamshaper8 checkpoint) · paper: Hessian Geometry of Latent Space in Generative Models