Formalizing the Sampling Design Space of Diffusion-Based Generative Models via Adaptive Solvers and Wasserstein-Bounded Timesteps
The episode discusses a paper formalizing sampling design space for diffusion models using adaptive solvers and Wasserstein-bounded timesteps. The hosts explain how this framework improves generation quality and efficiency by adapting solver selection based on trajectory geometry and scheduling using mathematical bounds, moving sampling from empirical art to a formal science.