Class DiffusionNoiseHelper<T>

Namespace

AiDotNet.Helpers

Assembly

AiDotNet.dll

Helper class for noise sampling operations in diffusion models.

public static class DiffusionNoiseHelper<T>

Type Parameters

T

The numeric type used for calculations.

Inheritance

object

DiffusionNoiseHelper<T>

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

This static helper provides common noise sampling operations used throughout diffusion models, ensuring consistent implementations and avoiding code duplication.

For Beginners: Diffusion models work by adding and removing noise from data. This helper provides the mathematical operations needed for that process: - Sampling Gaussian (bell-curve) noise - Computing noise schedules - Scaling noise for different timesteps

Methods

AddNoise(Tensor<T>, Tensor<T>, T, T)

Adds noise to a signal at a specified timestep using the scheduler's noise schedule.

public static Tensor<T> AddNoise(Tensor<T> signal, Tensor<T> noise, T sqrtAlphaCumprod, T sqrtOneMinusAlphaCumprod)

Parameters

signal Tensor<T>

The clean signal.

noise Tensor<T>

The noise to add.

sqrtAlphaCumprod T

The square root of cumulative alpha at this timestep.

sqrtOneMinusAlphaCumprod T

The square root of (1 - cumulative alpha) at this timestep.

Returns

Tensor<T>

The noisy signal: sqrt(alpha_cumprod) * signal + sqrt(1 - alpha_cumprod) * noise.

Remarks

For Beginners: This is how we add noise during training: - At timestep 0: Almost no noise (mostly signal) - At timestep T: Almost all noise (almost no signal) - The alphas control this blend based on timestep

ComputeSNR(T)

Computes the signal-to-noise ratio (SNR) for a given timestep.

public static T ComputeSNR(T alphaCumprod)

Parameters

alphaCumprod T

The cumulative alpha product at this timestep.

Returns

T

The SNR value: alpha / (1 - alpha).

ComputeTimestepEmbedding(int, int)

Computes sinusoidal embedding for a single timestep.

public static Vector<T> ComputeTimestepEmbedding(int timestep, int embeddingDim)

Parameters

timestep int

The timestep to embed.

embeddingDim int

The dimension of the embedding.

Returns

Vector<T>

Vector of length embeddingDim containing the embedding.

ComputeTimestepEmbeddings(int[], int)

Computes sinusoidal timestep embeddings (like in Transformers).

public static Tensor<T> ComputeTimestepEmbeddings(int[] timesteps, int embeddingDim)

Parameters

timesteps int[]

Array of timesteps to embed.

embeddingDim int

The dimension of each embedding.

Returns

Tensor<T>

Tensor of shape [batchSize, embeddingDim] containing the embeddings.

Remarks

For Beginners: This converts timestep numbers (like 100, 500, 999) into high-dimensional vectors that the neural network can understand. The sinusoidal pattern helps the network distinguish between nearby and distant timesteps.

LerpNoise(Tensor<T>, Tensor<T>, double)

Linearly interpolates between two noise tensors.

public static Tensor<T> LerpNoise(Tensor<T> noise1, Tensor<T> noise2, double t)

Parameters

noise1 Tensor<T>

First noise tensor.

noise2 Tensor<T>

Second noise tensor.

t double

Interpolation factor (0 = noise1, 1 = noise2).

Returns

Tensor<T>

Interpolated noise tensor.

SampleGaussian(int[], int?)

Samples Gaussian noise from a standard normal distribution N(0, 1).

public static Tensor<T> SampleGaussian(int[] shape, int? seed = null)

Parameters

shape int[]

The shape of the noise tensor to generate.

seed int?

Optional random seed for reproducibility.

Returns

Tensor<T>

A tensor filled with Gaussian noise.

Remarks

Uses the Box-Muller transform to convert uniform random numbers to Gaussian. The existing RandomHelper is used for thread-safe random number generation.

For Beginners: This creates random "static" like you might see on an old TV. Each value is drawn from a bell curve (Gaussian distribution) centered at 0. Most values will be close to 0, with occasional larger positive or negative values.

SampleGaussian(int[], Random)

Samples Gaussian noise using a provided random number generator.

public static Tensor<T> SampleGaussian(int[] shape, Random rng)

Parameters

shape int[]

The shape of the noise tensor to generate.

rng Random

The random number generator to use.

Returns

Tensor<T>

A tensor filled with Gaussian noise.

SampleGaussianVector(int, int?)

Samples Gaussian noise as a Vector.

public static Vector<T> SampleGaussianVector(int length, int? seed = null)

Parameters

length int

The length of the vector to generate.

seed int?

Optional random seed for reproducibility.

Returns

Vector<T>

A vector filled with Gaussian noise.

SampleGaussianVector(int, Random)

Samples Gaussian noise as a Vector using a provided random number generator.

public static Vector<T> SampleGaussianVector(int length, Random rng)

Parameters

length int

The length of the vector to generate.

rng Random

The random number generator to use.

Returns

Vector<T>

A vector filled with Gaussian noise.

ScaleNoise(Tensor<T>, double)

Scales noise by a given factor.

public static Tensor<T> ScaleNoise(Tensor<T> noise, double scale)

Parameters

noise Tensor<T>

The noise tensor to scale.

scale double

The scaling factor.

Returns

Tensor<T>

Scaled noise tensor.

SlerpNoise(Tensor<T>, Tensor<T>, double)

Spherical linear interpolation between two noise tensors.

public static Tensor<T> SlerpNoise(Tensor<T> noise1, Tensor<T> noise2, double t)

Parameters

noise1 Tensor<T>

First noise tensor.

noise2 Tensor<T>

Second noise tensor.

t double

Interpolation factor (0 = noise1, 1 = noise2).

Returns

Tensor<T>

Spherically interpolated noise tensor.

Remarks

For Beginners: SLERP is like traveling on the surface of a sphere instead of cutting through it. For noise interpolation, this often gives smoother transitions between different random seeds.