PixArtDeltaModel<T>

Models & Types · AiDotNet.Diffusion.TextToImage

PixArt-Delta model — LCM-distilled PixArt for fast 2-8 step generation.

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

For Beginners

PixArt-Delta is PixArt-Alpha made faster:

Key characteristics:

  • Same DiT architecture as PixArt-Alpha
  • LCM distillation: trained to produce good results in 2-8 steps
  • No classifier-free guidance needed (saves 50% compute)
  • Maintains T5-XXL text encoder for good prompt understanding

How it compares:

  • PixArt-Alpha: ~20 steps, needs CFG → slower
  • PixArt-Delta: ~4 steps, no CFG needed → 10x faster
  • SD Turbo: 1-4 steps but lower quality
  • PixArt-Delta: 2-8 steps with higher quality

Use PixArt-Delta when you need:

  • Fast generation (2-8 steps)
  • Good quality without guidance overhead
  • Resource-efficient deployment

How It Works

PixArt-Delta applies Latent Consistency Model (LCM) distillation to PixArt-Alpha, enabling high-quality image generation in just 2-8 denoising steps while preserving the efficient DiT architecture.

Technical specifications:

  • Architecture: DiT-XL/2 (same as PixArt-Alpha)
  • Distillation: Latent Consistency Model (LCM)
  • Steps: 2-8 (optimal: 4)
  • Guidance scale: 1.0 (no CFG)
  • Text encoder: T5-XXL (4096-dim)
  • Resolution: 512x512 to 1024x1024

Reference: Chen et al., “PixArt-delta: Fast and Controllable Image Generation with Latent Consistency Models”, 2024

Example

var options = new DiffusionModelOptions<float> { LatentChannels = 4, DefaultInferenceSteps = 20 };
var noise = Tensor<float>.CreateRandom(new[] { 1, 4, 128, 128 });
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 PixArtDeltaModel<float>(options: options))
    .Build(trainX, trainY);
var generated = result.Predict(noise);

Constructors

Constructor Summary
PixArtDeltaModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, DiTNoisePredictor<T> dit, StandardVAE<T> vae, IConditioningModule<T> conditioner, int? seed) Initializes a new instance of PixArtDeltaModel with full customization support.

Properties

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

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> GenerateFromText(string prompt, string negativePrompt, int width, int height, int numInferenceSteps, double? guidanceScale, int? seed)

Fields

Field Summary
int DefaultHeight
int DefaultWidth