INoiseSchedulerInputScaling<T>

Interfaces · AiDotNet.Interfaces

Optional sampler capability for schedules whose latent coordinate system differs from the denoiser’s model-input coordinate system, such as sigma-space Euler sampling.

Known implementations: EulerDiscreteScheduler<T>

Properties

Property Summary
T InitialNoiseSigma { get; } Scale applied to unit Gaussian initial noise after inference timesteps are configured.

Methods

Method Summary
Tensor<T> ScaleModelInput(Tensor<T> sample, int timestep) Converts only the evolving noisy latent to the denoiser’s input coordinates without mutating the sample.