PixArtSigmaModel<T>

Models & Types · AiDotNet.Diffusion.TextToImage

PixArt-Sigma model for high-resolution text-to-image generation with improved quality.

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

For Beginners

PixArt-Sigma is an upgraded version of PixArt-Alpha:

Key improvements over PixArt-Alpha:

  • Higher native resolution: supports up to 4096x4096
  • Better image quality from improved training data
  • Flexible aspect ratios via bucket training
  • Still very efficient (DiT-based, much faster than U-Net models)

How PixArt-Sigma works:

  1. Text goes through T5-XXL encoder (4096-dim)
  2. DiT transformer blocks denoise the latent with cross-attention to text
  3. Trained on high-quality curated datasets with better captions
  4. VAE decodes to final high-resolution image

Use PixArt-Sigma when you need:

  • High-resolution output (up to 4K)
  • Fast generation on limited hardware
  • Good quality with flexible aspect ratios

How It Works

PixArt-Sigma is the successor to PixArt-Alpha, featuring improved training data quality, higher native resolution support (up to 4K), and better aesthetic quality. It maintains the efficient DiT architecture while achieving quality comparable to SDXL and DALL-E 3.

Technical specifications:

  • Architecture: DiT (Diffusion Transformer) with T5-XXL text encoder
  • Parameters: ~600M (DiT-XL/2)
  • Text encoder: T5-XXL (4096-dim embeddings)
  • Native resolution: up to 4096x4096
  • Latent space: 4 channels, 8x downsampling
  • Scheduler: DPM-Solver with 20 steps recommended

Reference: Chen et al., “PixArt-Sigma: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation”, ECCV 2024

Example

var options = new DiffusionModelOptions<float> { LatentChannels = 4, DefaultInferenceSteps = 20 };
var noise = Tensor<float>.CreateRandom(new[] { 1, 4, 256, 256 });
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 PixArtSigmaModel<float>(options: options))
    .Build(trainX, trainY);
var generated = result.Predict(noise);

Constructors

Constructor Summary
PixArtSigmaModel(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 PixArtSigmaModel 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