The Noise Schedule: Beta Values
You'll learn: How to control the speed and smoothness of noise addition in diffusion models using variance schedules defined by beta values.
What Are Beta Values?
In the forward diffusion process, we gradually add noise to images over many timesteps (often 1000). But how much noise do we add at each step? That's controlled by beta values (β_t) — small numbers that define the variance of noise at timestep t.
Think of beta values like the "intensity dial" at each step of your noise-adding journey. Small betas add tiny amounts of noise; larger betas add more. The complete sequence of beta values across all timesteps is called the noise schedule.
Why Schedules Matter
You can't just pick random beta values. The schedule should:
- Start with small betas (preserve image structure early on)
- End with larger betas (approach pure noise at the end)
- Transition smoothly to help the model learn denoising
Common Variance Schedules
Linear Schedule
The simplest approach: beta increases linearly from a small start value (e.g., 0.0001) to a larger end value (e.g., 0.02). Easy to implement but can be suboptimal — it might add too much noise too quickly in the middle timesteps.
Cosine Schedule
A more refined approach that uses a cosine curve to determine beta values. This keeps smaller changes at the beginning and end, with smoother transitions. Often produces better sample quality because it maintains signal longer.
The choice of schedule dramatically affects how well your model learns to denoise and generate images.
Key Takeaway: Beta values define your noise schedule — the recipe for how aggressively noise is added at each diffusion timestep, with linear and cosine schedules being popular choices.