Imagen3Model<T>
Models & Types · AiDotNet.Diffusion.TextToImage
Imagen 3 model for text-to-image generation by Google DeepMind.
Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → Imagen3Model<T>
For Beginners
Imagen 3 is Google DeepMind’s latest image generation model.
How Imagen 3 works:
- Text is encoded by a Gemma text encoder (evolved from T5-XXL)
- A base SiT (Scalable Interpolant Transformer) generates a 64x64 latent
- A super-resolution cascade upscales to 256x256, then 1024x1024
- Each stage is a separate diffusion model operating at increasing resolution
- Human preference alignment ensures quality and safety
Key characteristics:
- Cascaded architecture: base (64x64) + SR (256x256) + SR (1024x1024)
- Gemma text encoder for deep prompt understanding
- Human feedback alignment (RLHF-style) for quality and safety
- SiT-based backbone with interpolant formulation
- 16 latent channels
Advantages:
- Exceptional prompt adherence and photorealism
- Strong safety filtering via RLHF alignment
- Excellent text rendering capabilities
- Cascaded approach enables very high resolution
Limitations:
- API-only (not open-source)
- Cascaded pipeline is slower than single-stage models
- Higher compute requirements due to multiple stages
How It Works
Imagen 3 uses a cascaded diffusion approach with a base model and super-resolution stages. It features Gemma-based text understanding (evolved from T5) and is aligned with human feedback through RLHF-style training for improved safety and aesthetic quality.
Technical specifications:
- Architecture: Cascaded SiT with super-resolution stages
- Base: ~2B params, SiT backbone, 64x64 latent generation
- SR Stage 1: 256x256 upscale
- SR Stage 2: 1024x1024 final output
- Text encoder: Gemma (2048-dim embeddings)
- Latent channels: 16
- Default: 50 steps, guidance scale 7.5
- Resolution: 1024x1024 default, up to 2048x2048
Reference: Google DeepMind, “Imagen 3”, 2024
Example
// Create Imagen 3
var imagen3 = new Imagen3Model<float>();
// Generate a photorealistic image
var image = imagen3.GenerateFromText(
prompt: "A professional photo of a golden retriever playing in autumn leaves",
width: 1024,
height: 1024,
numInferenceSteps: 50,
guidanceScale: 7.5,
seed: 42);
Constructors
| Constructor | Summary |
|---|---|
Imagen3Model(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, SiTPredictor<T> predictor, StandardVAE<T> vae, IConditioningModule<T> conditioner, int? seed) |
Initializes a new instance of Imagen3Model. |
Properties
| Property | Summary |
|---|---|
IConditioningModule<T> Conditioner { get; } |
|
INoisePredictor<T> NoisePredictor { get; } |
|
IVAEModel<T> VAE { get; } |
|
bool UsesCascadedArchitecture { get; } |
Gets whether this model uses a cascaded architecture with super-resolution stages. |
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) |
|
Tensor<T> ImageToImage(Tensor<T> inputImage, string prompt, string negativePrompt, double strength, int numInferenceSteps, double? guidanceScale, int? seed) |
Fields
| Field | Summary |
|---|---|
int DefaultHeight |
Default image height for Imagen 3 (1024x1024). |
int DefaultWidth |
Default image width for Imagen 3 (1024x1024). |