Classifier-Free Guidance - From High-Dimensional Analysis to Generalized Guidance Forms

Explore the mathematical foundations and practical applications of classifier-free guidance in this 46-minute research presentation by Krunoslav Lehman Pavasovic from the Generative Memory Lab. Delve into high-dimensional analysis techniques used to understand how classifier-free guidance operates in generative models, examining the theoretical underpinnings that make this approach effective for controlling generation processes without requiring separate classifier networks. Learn about generalized guidance forms that extend beyond traditional classifier-free guidance, discovering new mathematical frameworks that can be applied to various generative modeling scenarios. Gain insights into the latest research developments in guidance mechanisms for diffusion models and other generative architectures, with detailed explanations of the analytical methods used to study these systems in high-dimensional spaces. Access the accompanying research paper at https://arxiv.org/abs/2502.07849 to follow along with the mathematical derivations and experimental results discussed throughout the presentation.