RAPHAELModel<T>
Models & Types · AiDotNet.Diffusion.TextToImage
RAPHAEL model — Mixture-of-Experts text-to-image diffusion model.
Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → RAPHAELModel<T>
For Beginners
RAPHAEL uses specialized experts for different concepts:
How RAPHAEL works:
- Text prompt is parsed into concept tokens
- Each concept routes to specialized expert pathways in cross-attention
- Space-MoE: different spatial regions use different experts
- Time-MoE: different denoising steps use different experts
- This enables precise control over each concept’s visual appearance
Key characteristics:
- MoE cross-attention: each token gets specialized experts
- Space-MoE: experts specialize in spatial regions
- Time-MoE: experts specialize in denoising stages
- Better text-image alignment than standard U-Net
Use RAPHAEL when you need:
- Precise text-to-image alignment
- Multi-concept scenes with distinct attributes
- High aesthetic quality
How It Works
RAPHAEL uses a novel Mixture-of-Experts (MoE) approach in the cross-attention layers of a diffusion U-Net. Each text concept activates different expert pathways, enabling fine-grained alignment between text tokens and image regions.
Technical specifications:
- Architecture: U-Net with MoE cross-attention layers
- Expert routing: Top-k routing per token/region/timestep
- Text encoder: CLIP ViT-L + edge-of-concept attention
- Resolution: 1024x1024
- Latent channels: 4
Reference: Xue et al., “RAPHAEL: Text-to-Image Generation via Large Mixture of Diffusion Paths”, NeurIPS 2023
Example
var options = new DiffusionModelOptions<float> { LatentChannels = 4, DefaultInferenceSteps = 30 };
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 RAPHAELModel<float>(options: options))
.Build(trainX, trainY);
var generated = result.Predict(noise);
Constructors
| Constructor | Summary |
|---|---|
RAPHAELModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, UNetNoisePredictor<T> unet, StandardVAE<T> vae, IConditioningModule<T> conditioner, int? seed) |
Initializes a new instance of RAPHAELModel 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 |