Geometric Asymptotics of Score Mixing and Guidance in Diffusion Models

Abstract

Diffusion models are routinely guided in practice by combining multiple score fields, yet the mathematical structure of score mixing is still poorly understood. We study the small-time generation dynamics driven by mixed scores $$ s=λ\,\nabla\log u_1+(1-λ)\,\nabla\log u_2,\qquad λ\ge 0, $$ in the heat-flow framework, where $u_1,u_2$ are heat evolutions of two compactly supported probability measures. This single formulation covers both the mixture-of-experts regime $(0\leq λ\leq 1)$ and the classifier-free guidance regime $(λ>1)$. Exploiting a Laplace-Varadhan principle under a similarity-time rescaling, we show that the small-time generation dynamics is governed by the explicit geometric potential $$ Φ_λ=λd_1^2+(1-λ)d_2^2, $$ which depends only on the supports of the initial measures and on the mixing parameter. This gives a rigorous reduction from a singular, non-autonomous score-driven dynamics to autonomous Clarke-type subgradient inclusions. In the empirical setting of finite Dirac mixtures, the limiting potential is piecewise quadratic with a Voronoi-type structure; this rigidity yields convergence of all autonomous limiting trajectories to critical points and a conditional convergence criterion for the original generation flow toward local minimizers of the potential, with rate $\mathcal O(\sqrt t)$ in the smooth stable case.