Course contentsShow
← Machine Learning and Deep Learning
Lesson 1526 of 3,53834. Diffusion ModelsFree lesson

Variance Schedule: Controlling Noise Addition

Learn how beta schedules control the amount of noise added at each diffusion step.

Variance Schedule: Controlling Noise Addition

What you'll learn: How beta schedules precisely control how much noise gets added at each step of the diffusion process.

The Core Idea

In the forward diffusion process, we gradually add noise to an image over many timesteps. But not all timesteps should add the same amount of noise! The variance schedule (also called the beta schedule) is a sequence of values β₁, β₂, ..., βₜ that determines exactly how much noise to add at each step.

Think of it like adjusting the volume on a radio: you wouldn't jump from silence to maximum volume instantly. Instead, you'd turn the dial gradually. The variance schedule is your "volume control" for noise.

Why It Matters

The choice of variance schedule dramatically affects:

  • Training stability: Too much noise too fast can destroy information prematurely
  • Sample quality: A good schedule preserves image structure in early steps
  • Computational efficiency: Well-designed schedules can use fewer timesteps

Common Variance Schedules

Linear Schedule

The simplest approach: β increases linearly from β₁ (e.g., 0.0001) to βₜ (e.g., 0.02). Small noise additions early, larger ones later.

Cosine Schedule

Uses a cosine function to add noise more smoothly, avoiding sudden jumps. This often produces better results by maintaining more signal longer.

The schedule essentially answers: "At timestep t, what fraction of the image should be replaced with noise?" A good schedule starts conservative and gradually increases destruction.

Key Takeaway: The variance schedule (β values) is the "knob" that controls how aggressively noise is added at each diffusion step—getting this right is crucial for high-quality generation.