Cosine Schedule

The cosine schedule was designed to fix the linear schedule's tendency to destroy information too quickly. The cosine-shaped $\bar\alpha_t$ curve provides a more gradual noise progression: the rate of change is highest near the midpoint (resembling a linear drop-off), but smooths out near $t = 0$ and $t = T$, ensuring that only minimal noise is added at the very beginning and very end. This delays the hardest denoising tasks until after the midpoint, leading to improved sample quality, better training efficiency, and faster convergence compared to the linear schedule.