Recent advancements in Diffusion models (DMs) have explored incorporating the second-order diffusion Fisher information (DF) into various downstream tasks and theoretical analysis.
Diffusion Fisher can be accessed efficiently within a space spanned by the outer products of score and initial data, leading to two efficient approximation algorithms for trace and matrix-vector multiplication of DF.
These algorithms bypass the auto-differentiation operations and provide time-efficient vector-product calculations, with established approximation error bounds.
Experiments show superior accuracy and reduced computational cost of the proposed algorithms, along with the design of a novel numerical verification experiment for the optimal transport property of the general PF-ODE deduced map.