Adaptive Reparametrized Time for Score-Based Diffusion Sampling
The episode discusses 'Adaptive Reparametrized Time' (ART), a continuous-time control formulation for score-based diffusion sampling that learns optimal timestep allocation. The hosts explain how ART treats time progression as a controllable rate to adapt resolution along the reverse trajectory based on numerical error sensitivities, leading to improved performance and transferability across different models and datasets.