Enum BetaSchedule

Namespace

AiDotNet.Enums

Assembly

AiDotNet.dll

Defines the types of beta (noise variance) schedules available for diffusion models.

public enum BetaSchedule

Fields

Linear = 0

Linear interpolation between beta start and end values.

The simplest schedule: beta increases linearly from start to end. This is the original schedule used in the DDPM paper.

Default values: beta_start=0.0001, beta_end=0.02 (from DDPM paper)

ScaledLinear = 1

Scaled linear schedule commonly used in latent diffusion models.

Uses the square root of linearly interpolated values. This is the default schedule used by Stable Diffusion and similar models.

Default values: beta_start=0.00085, beta_end=0.012 (from Stable Diffusion)

SquaredCosine = 2

Squared cosine schedule for improved diffusion models.

Based on the "Improved Denoising Diffusion Probabilistic Models" paper. Provides smoother noise progression and often better generation quality.

Why use this: The squared cosine schedule is designed to maintain more signal (clearer images) for longer during the forward process, which can improve the model's ability to learn fine details.

Remarks

The beta schedule controls how noise variance changes across timesteps during the diffusion process. Different schedules have different characteristics and are suited for different applications.

For Beginners: Think of this like choosing how to gradually add static to a TV signal.

• Linear: Add static evenly - each step adds about the same amount
• ScaledLinear: Start slow, then add more - common in image generation (Stable Diffusion)
• SquaredCosine: Smooth S-curve - often produces better quality results

The choice affects both training efficiency and generation quality.