SelfAttentionGuidance<T>

Models & Types · AiDotNet.Diffusion.Guidance

Self-Attention Guidance (SAG) for diffusion model inference.

Implements: IGuidanceMethod<T>

For Beginners

SAG looks at which parts of the image the model pays most attention to, then uses that information to guide generation more precisely. It can improve detail in important areas while reducing artifacts.

How It Works

SAG leverages self-attention maps to selectively blur high-attention regions, creating an intermediate prediction between conditional and unconditional. This provides more focused guidance than standard CFG.

Reference: Hong et al., “Improving Sample Quality of Diffusion Models Using Self-Attention Guidance”, ICCV 2023

Constructors

Constructor Summary
SelfAttentionGuidance(double sagScale, double blurSigma) Initializes a new Self-Attention Guidance instance.

Properties

Property Summary
GuidanceType GuidanceType { get; }

Methods

Method Summary
Tensor<T> Apply(Tensor<T> unconditional, Tensor<T> conditional, double scale, double timestep)