Browse Teach Curate

Optimizing Noise Schedules of Diffusion Models in High Dimensions

Within the stochastic-interpolants framework, tailoring the variance-preserving (VP) and variance-exploding (VE) noise schedules so a diffusion model's probability-flow ODE recovers both low-level (per-mode) and high-level (between-mode asymmetry) structure and resolves the dimension-dependent speciation transition, reducing the ODE discretization cost from Theta_d(sqrt(d)) steps under constant denoising to Theta_d(1) steps in dimension d.

Data Text

Optimizing Noise Schedules of Generative Models in High Dimensions Santiago Aranguri Giulio Biroli

A theoretical machine-learning / statistical-physics study of the sampling error of diffusion-based generative models in the high-dimensional limit, cast in the stochastic-interpolants framework and …