Class SchedulerConfig<T>

Namespace

AiDotNet.NeuralNetworks.Diffusion.Schedulers

Assembly

AiDotNet.dll

Configuration options for diffusion model step schedulers.

public sealed class SchedulerConfig<T>

Type Parameters

T

The numeric type used for calculations.

Inheritance

object

SchedulerConfig<T>

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

This configuration class defines all the parameters needed to initialize a step scheduler. These parameters control the noise schedule and behavior of the diffusion process.

For Beginners: This is like a settings panel for the scheduler. You can control:

• How many steps to use (more = higher quality, slower)
• How much noise to start and end with (the beta values)
• What pattern of noise to use (linear, scaled, cosine)
• Whether to clip values to prevent extreme outputs
• What the model is predicting (noise, sample, or velocity)

The default values are research-backed and work well for most cases.

Constructors

SchedulerConfig(int, T, T, BetaSchedule, bool, DiffusionPredictionType)

Initializes a new scheduler configuration.

public SchedulerConfig(int trainTimesteps, T betaStart, T betaEnd, BetaSchedule betaSchedule = BetaSchedule.Linear, bool clipSample = false, DiffusionPredictionType predictionType = DiffusionPredictionType.Epsilon)

Parameters

trainTimesteps int

Number of timesteps for training. Must be greater than 1. Default: 1000.

betaStart T

Starting beta value. Must be positive. Default: 0.0001.

betaEnd T

Ending beta value. Must be greater than betaStart. Default: 0.02.

betaSchedule BetaSchedule

The beta schedule type. Default: Linear.

clipSample bool

Whether to clip samples to [-1, 1]. Default: false.

predictionType DiffusionPredictionType

What the model predicts. Default: Epsilon.

Exceptions

ArgumentOutOfRangeException

Thrown when trainTimesteps is less than 2, or when beta values are invalid.

Properties

BetaEnd

Gets the ending beta value (noise variance at t=T).

public T BetaEnd { get; }

Property Value

T

Remarks

Beta end is the amount of noise added at the final timestep. It should be larger than beta start to ensure samples become nearly pure noise.

Default: 0.02 for Linear schedule (from DDPM paper)

BetaSchedule

Gets the type of beta schedule to use.

public BetaSchedule BetaSchedule { get; }

Property Value

BetaSchedule

Remarks

Different schedules provide different noise progression characteristics. Linear is simplest, ScaledLinear is used by Stable Diffusion, SquaredCosine often provides better quality.

Default: Linear (most widely compatible)

BetaStart

Gets the starting beta value (noise variance at t=0).

public T BetaStart { get; }

Property Value

T

Remarks

Beta start is the amount of noise added at the first timestep. It should be small since early timesteps should preserve most of the signal.

Default: 0.0001 for Linear schedule (from DDPM paper)

ClipSample

Gets whether to clip predicted samples to [-1, 1].

public bool ClipSample { get; }

Property Value

bool

Remarks

Clipping can prevent extreme values that might cause numerical instability, but may reduce output quality if the model is well-trained.

Default: false (let the model produce natural outputs)

PredictionType

Gets the type of prediction the model makes.

public DiffusionPredictionType PredictionType { get; }

Property Value

DiffusionPredictionType

Remarks

This must match what the diffusion model was trained to predict. Most models use Epsilon (noise) prediction.

Default: Epsilon (most common)

TrainTimesteps

Gets the number of timesteps used during training.

public int TrainTimesteps { get; }

Property Value

int

Remarks

This is typically 1000 for most diffusion models. More timesteps allow for finer-grained noise schedules but increase training time.

Default: 1000 (standard for DDPM/DDIM)

Methods

CreateDefault()

Creates a default configuration for DDPM-style models.

public static SchedulerConfig<T> CreateDefault()

Returns

SchedulerConfig<T>

A scheduler configuration with standard DDPM defaults.

Remarks

Uses values from the original DDPM paper: - 1000 training timesteps - Beta start: 0.0001 - Beta end: 0.02 - Linear beta schedule - Epsilon prediction

CreateStableDiffusion()

Creates a configuration optimized for Stable Diffusion-style models.

public static SchedulerConfig<T> CreateStableDiffusion()

Returns

SchedulerConfig<T>

A scheduler configuration with Stable Diffusion defaults.

Remarks

Uses values from Stable Diffusion: - 1000 training timesteps - Beta start: 0.00085 - Beta end: 0.012 - Scaled linear beta schedule - Epsilon prediction