LatentInitializer<T>

Models & Types ยท AiDotNet.Diffusion.Schedulers

Initializes latent tensors for diffusion generation with various noise strategies.

For Beginners

Before a diffusion model can create an image, it needs a starting point โ€” usually pure random noise. LatentInitializer handles creating this starting noise. For text-to-image, it creates pure noise. For image editing (img2img), it mixes the original image with noise, controlling how much of the original to keep.

How It Works

Provides different strategies for initializing the starting noise latent for diffusion generation. Supports standard Gaussian noise, image-conditioned initialization (img2img), and strength-based partial noise for editing workflows. Ensures consistent shapes and proper scaling for the target scheduler.

Constructors

Constructor Summary
LatentInitializer(double initNoiseSigma, int? seed) Initializes a new latent initializer.

Properties

Property Summary
double InitNoiseSigma { get; } Gets the initial noise sigma scaling factor.

Methods

Method Summary
Vector<T> CreateImg2ImgLatent(Vector<T> encodedImage, double strength) Creates a latent initialized from an encoded image with noise added (img2img).
Vector<T> CreateNoiseLatent(int size) Creates a pure noise latent for text-to-image generation.