Persistent Representation Learning for Drifting Models
October 2, 2026
Persistent representation learning improves one-step generation in Drifting Models by continuously learning a discriminative representation geometry as the generator evolves. This addresses the performance gap in pixel-space drifting by optimizing the sample weighting used in kernel density estimation.
HOW THIS AFFECTS YOU
●
researcherYou can improve the sample quality of one-step generative models by stabilizing the underlying representation geometry.