AdversarialDistillationTrainer<T>

Models & Types · AiDotNet.Diffusion.Distillation

Trainer for Adversarial Diffusion Distillation (ADD) as used in SD/SDXL Turbo.

For Beginners

ADD uses two teachers: (1) a diffusion model that ensures the student captures the right image structure, and (2) a discriminator that ensures the output looks realistic. This dual supervision produces remarkably good single-step images — the student learns both “what” to generate and “how real” it should look.

How It Works

ADD combines a diffusion denoising loss with an adversarial loss from a pretrained discriminator. The student generates a sample in 1-4 steps, a discriminator evaluates realism, and the diffusion teacher provides structure guidance. This dual-loss approach produces the highest quality single-step generators.

Reference: Sauer et al., “Adversarial Diffusion Distillation”, 2023

Constructors

Constructor Summary
AdversarialDistillationTrainer(IDiffusionModel<T> teacher, double adversarialWeight, double diffusionWeight) Initializes a new adversarial distillation trainer.

Properties

Property Summary
IDiffusionModel<T> Teacher { get; } Gets the teacher model.
double AdversarialWeight { get; } Gets the adversarial loss weight.
double DiffusionWeight { get; } Gets the diffusion loss weight.

Methods

Method Summary
T ComputeADDLoss(Vector<T> studentOutput, Vector<T> teacherOutput, T discriminatorScore) Computes the combined ADD loss.