IAIFI Summer Workshop Talk

Diffusion models generate complex data by estimating the score—the gradient of the log-density—across varying noise scales, but the relationship between the learned neural score and the true data score has remained unclear. For moderate-to-high noise levels, the learned score is dominated by its linear (Gaussian) component, enabling a closed-form integration of the probability-flow ODE. This analytical solution predicts key sampling phenomena including the early specification of coarse structures (e.g., scene layouts), the low dimensionality of sampling trajectories, and their sensitivity to perturbations.

Aug 11, 2025 12:00 AM — Aug 15, 2025 12:00 AM