Generative Drifting is Secretly Score Matching: a Spectral and Variational Perspective
A new paper proposes that Generative Drifting, a method for one-step image generation, is fundamentally a form of score matching. The research reveals that under specific conditions, the drift operator in this technique is equivalent to calculating score differences on smoothed distributions. This insight helps explain the necessity of the stop-gradient operator for stable training and suggests optimizations for kernel selection and convergence speed, drawing parallels to plasma physics. AI
IMPACT Provides a theoretical framework for generative models, potentially leading to more efficient and stable training methods.