LaplaceNoiseSchedule<T>
Models & Types · AiDotNet.Diffusion.Schedulers.NoiseSchedules
Laplace noise schedule using heavy-tailed Laplace distribution for noise sampling.
For Beginners
Laplace noise has more “extreme” values than standard Gaussian noise. Using it during training helps the model handle sharp edges and high-contrast areas better, leading to crisper generated images.
How It Works
Replaces Gaussian noise with Laplace-distributed noise during the forward process. The heavier tails of the Laplace distribution help the model handle extreme values better, improving generation of high-contrast and high-frequency details.
Constructors
| Constructor | Summary |
|---|---|
LaplaceNoiseSchedule(double scale) |
Initializes a new instance with the specified Laplace scale parameter. |
Methods
| Method | Summary |
|---|---|
Vector<T> SampleNoise(int length, Random random) |
Generates Laplace-distributed noise values. |