<ul data-eligibleForWebStory="true"><li>Diffusion-based generative models use SDEs and probability flow ODEs to transform complex data distributions into manageable prior distributions.</li><li>A geometric regularity has been discovered in the deterministic sampling dynamics of these models, where trajectories exhibit a consistent 'boomerang' shape in an extremely low-dimensional subspace.</li><li>This regularity persists regardless of model architecture, conditions, or generated content and has been studied using kernel-estimated data modeling.</li><li>The research also introduces a dynamic programming-based scheme to synchronize the sampling time schedule with the identified trajectory structure, leading to enhanced image generation performance with minimal computational overhead.</li></ul>

Geometric Regularity in Deterministic Sampling of Diffusion-based Generative Models

Discover more