ImprovedConsistencyModel<T>

Models & Types · AiDotNet.Diffusion.FastGeneration

Improved Consistency Training (iCT) model for single-step image generation.

Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → ImprovedConsistencyModel<T>

For Beginners

Consistency models learn to map any noisy image directly to the clean image in a single step. Unlike diffusion models that remove noise gradually over 20-50 steps, iCT does it in one shot. The “improved” version trains better by using smarter loss functions and noise schedules, producing higher quality single-step images.

How It Works

iCT improves upon the original Consistency Training by using a lognormal schedule for the noise discretization, pseudo-Huber loss instead of L2, and an improved EMA schedule. Achieves state-of-the-art FID scores among single-step generators on ImageNet 64x64.

Reference: Song and Dhariwal, “Improved Techniques for Training Consistency Models”, ICLR 2024

Example

var options = new DiffusionModelOptions<float> { LatentChannels = 4, DefaultInferenceSteps = 1 };
var noise = Tensor<float>.CreateRandom(new[] { 1, 4, 8, 8 });
var trainX = Tensor<float>.CreateRandom(4, 8);
var trainY = Tensor<float>.CreateRandom(4, 2);
var result = new AiModelBuilder<float, Tensor<float>, Tensor<float>>()
    .ConfigureModel(new ImprovedConsistencyModel<float>(options: options))
    .Build(trainX, trainY);
var generated = result.Predict(noise);

Constructors

Constructor Summary
ImprovedConsistencyModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, UNetNoisePredictor<T> predictor, StandardVAE<T> vae, IConditioningModule<T> conditioner, int? seed) Initializes a new Improved Consistency Training model.

Properties

Property Summary
IConditioningModule<T> Conditioner { get; }
INoisePredictor<T> NoisePredictor { get; }
IVAEModel<T> VAE { get; }
int LatentChannels { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()