OffsetNoiseSchedule<T>

Models & Types · AiDotNet.Diffusion.Schedulers.NoiseSchedules

Offset noise schedule that adds a global offset to noise for improved dark/bright image generation.

For Beginners

Standard diffusion models struggle to generate very dark or very bright images because the noise always averages out to medium brightness. Offset noise fixes this by occasionally adding brightness shifts to the entire image during training.

How It Works

Standard diffusion noise is zero-mean per pixel, which biases the model toward mid-tone images. Offset noise adds a small per-channel offset (same value for all pixels in a channel), enabling the model to generate very dark or very bright images.

Reference: Originally proposed by Nicholas Guttenberg, widely adopted in SD community

Constructors

Constructor Summary
OffsetNoiseSchedule(double offsetStrength) Initializes a new instance with the specified offset strength.

Methods

Method Summary
Vector<T> ApplyOffset(Vector<T> noise, int numChannels, Random random) Applies offset noise to a standard Gaussian noise vector.