Noise Schedule
The predefined schedule of noise levels added during diffusion model training and removed during generation.
Overview
A noise schedule defines how much Gaussian noise is added to data at each timestep during the forward diffusion process in diffusion models. It controls the rate at which data is corrupted to pure noise (forward process) and subsequently denoised back to clean data (reverse process).
Key Details
Common schedules include linear, cosine, and sigmoid schedules, each affecting generation quality differently. The cosine schedule (introduced with improved DDPM) produces better results by adding noise more gradually at the beginning and end. The noise schedule is a critical hyperparameter — it affects training stability, sample quality, and the number of diffusion steps needed for generation.