SDXLModel<T>

Models & Types · AiDotNet.Diffusion.TextToImage

Stable Diffusion XL (SDXL) model for high-resolution image generation.

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

For Beginners

SDXL is like Stable Diffusion 2.0 but significantly upgraded:

Key improvements over SD 1.5/2.0:

  • 4x larger U-Net (2.6B vs 865M parameters)
  • Dual text encoders (better prompt understanding)
  • Native 1024x1024 resolution (vs 512x512)
  • Optional refiner model for enhanced details

How SDXL works:

  1. Your prompt goes through TWO text encoders (CLIP + OpenCLIP)
  2. These embeddings guide a much larger U-Net during denoising
  3. The base model generates at 1024x1024
  4. (Optional) A refiner model enhances fine details

Example prompt flow: “A majestic dragon” -> [CLIP] + [OpenCLIP] -> Combined embedding -> Large U-Net denoises -> 1024x1024 image -> (Optional) Refiner -> Enhanced details

Use SDXL when you need:

  • High resolution output
  • Better text rendering in images
  • More detailed and coherent images
  • Following complex prompts accurately

How It Works

SDXL is Stability AI’s flagship text-to-image model, designed for high-quality 1024x1024 image generation with improved prompt understanding and visual fidelity compared to earlier Stable Diffusion versions.

Technical specifications:

  • Base model: 2.6B parameter U-Net
  • Text encoders: CLIP ViT-L/14 + OpenCLIP ViT-bigG/14
  • Native resolution: 1024x1024
  • Latent space: 4 channels, 8x spatial downsampling
  • Guidance scale: 5.0-9.0 recommended (7.5 default)
  • Scheduler: DDPM/DPM++/Euler with 20-50 steps

Architecture details:

  • Micro-conditioning: Size and crop coordinates for multi-aspect training
  • Dual text encoding: Concatenated CLIP + OpenCLIP embeddings
  • Channel multipliers: [1, 2, 4, 4] (vs [1, 2, 4, 8] in SD 2.x)
  • Cross-attention dimension: 2048 (vs 1024 in SD 1.x)

Example

// Create an SDXL model
var sdxl = new SDXLModel<float>();
            
// Generate a high-resolution image
var image = sdxl.GenerateFromText(
    prompt: "A majestic dragon perched on a mountain peak at sunset, highly detailed",
    negativePrompt: "blurry, low quality, distorted",
    width: 1024,
    height: 1024,
    numInferenceSteps: 30,
    guidanceScale: 7.5,
    seed: 42);
            
// Generate with micro-conditioning for aspect ratio
var wideImage = sdxl.GenerateWithMicroCondition(
    prompt: "Panoramic landscape with mountains and lake",
    width: 1536,
    height: 640,
    originalWidth: 1536,
    originalHeight: 640,
    cropTop: 0,
    cropLeft: 0);
            
// Use the refiner for enhanced details
if (sdxl.SupportsRefiner)
{
    var refined = sdxl.RefineImage(image, "enhance details");
}

Constructors

Constructor Summary
SDXLModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, UNetNoisePredictor<T> unet, StandardVAE<T> vae, IConditioningModule<T> conditioner1, IConditioningModule<T> conditioner2, SDXLRefiner<T> refiner, bool useDualEncoder, int crossAttentionDim, int? seed) Initializes a new instance of SDXLModel with full customization support.

Properties

Property Summary
IConditioningModule<T> Conditioner { get; }
IConditioningModule<T> SecondaryConditioner { get; } Gets the secondary text encoder if available.
INoisePredictor<T> NoisePredictor { get; }
IVAEModel<T> VAE { get; }
SDXLRefiner<T> Refiner { get; } Gets the refiner model if available.
bool SupportsRefiner { get; } Gets whether this model has a refiner available.
bool UsesDualEncoder { get; } Gets whether this model uses dual text encoders.
int CrossAttentionDim { get; } Gets the cross-attention dimension (2048 for SDXL).
int LatentChannels { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Task<Tensor<T>> GenerateAsync(string prompt, string negativePrompt, int width, int height, int? originalWidth, int? originalHeight, int cropTop, int cropLeft, int numInferenceSteps, double? guidanceScale, int? seed, IProgress<int> progress, CancellationToken cancellationToken) Async wrapper around GenerateWithMicroCondition with cooperative cancellation between scheduler steps and per-step progress reporting.
Tensor<T> GenerateWithMicroCondition(string prompt, string negativePrompt, int width, int height, int? originalWidth, int? originalHeight, int cropTop, int cropLeft, int numInferenceSteps, double? guidanceScale, int? seed) Generates an image with micro-conditioning for multi-aspect ratio support.
Tensor<T> RefineImage(Tensor<T> image, string prompt, string negativePrompt, int numInferenceSteps, double denoiseStrength, int? seed) Refines an image using the SDXL refiner model.
void InvalidateAllStageCompiledPlans() Drops every stage’s plan in lockstep.
void InvalidateConditioner1CompiledPlans() Drops the conditioner1 stage’s compiled plan on the next generation.
void InvalidateConditioner2CompiledPlans() Drops conditioner2’s plan only.
void InvalidateUNetCompiledPlans() Drops UNet’s per-step plan only.
void InvalidateVAECompiledPlans() Drops VAE-decoder’s plan only.

Fields

Field Summary
int DefaultHeight Default height for SDXL generation.
int DefaultWidth Default width for SDXL generation.