HiDreamModel<T>
Models & Types · AiDotNet.Diffusion.TextToImage
HiDream-I1 model for high-quality imaginative text-to-image generation.
Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → HiDreamModel<T>
For Beginners
HiDream is designed to generate more creative and imaginative images with excellent prompt understanding.
How HiDream works:
- Text is encoded by CLIP ViT-L/14 and Llama-3.1 (a powerful language model)
- An MMDiT-X transformer processes text and image tokens with joint attention
- Flow matching training enables efficient 28-50 step generation
- A 16-channel VAE decodes latents to high-resolution images
Model variants:
- HiDream-I1 Full: 17B parameters, highest quality
- HiDream-I1 Dev: 12B parameters, good balance of quality and speed
- HiDream-I1 Fast: 8B parameters, optimized for speed
Key characteristics:
- Llama-3.1 as text encoder for deep language understanding
- MMDiT-X architecture with enhanced cross-attention
- 16 latent channels
- Flow matching training
- Excellent at artistic and fantasy-style content
- Strong composition and spatial reasoning
Advantages:
- Superior prompt understanding via Llama-3.1
- Great artistic and creative generation
- Multiple speed/quality tradeoff variants
- Open-weight model
Limitations:
- Full variant requires significant VRAM (~40GB)
- Newer model with smaller community
- Llama-3.1 encoder adds overhead
How It Works
HiDream-I1 uses an MMDiT architecture enhanced with Llama-3.1 as a text encoder, providing superior prompt understanding through a large language model backbone. It features improved composition handling and artistic style diversity compared to standard diffusion models, with variants ranging from 8B to 17B parameters.
Technical specifications:
- Architecture: MMDiT-X with Llama-3.1 text conditioning
- Full: 17B params, hidden 4096, 40 layers
- Dev: 12B params, hidden 3072, 32 layers
- Fast: 8B params, hidden 2560, 24 layers
- Text encoder 1: CLIP ViT-L/14 (768-dim, pooled)
- Text encoder 2: Llama-3.1-8B (4096-dim, sequence)
- Patch size: 2
- VAE: 16 latent channels, 8x spatial compression
- Training: Flow matching
- Resolution: 1024x1024 default
Reference: HiDream.ai, “HiDream-I1: High-Quality Imaginative Image Generation”, 2025
Example
// Create HiDream Full (highest quality)
var hiDream = new HiDreamModel<float>(variant: HiDreamVariant.Full);
// Create HiDream Fast for quicker generation
var hiDreamFast = new HiDreamModel<float>(variant: HiDreamVariant.Fast);
// Generate an image
var image = hiDream.GenerateFromText(
prompt: "A mystical forest with bioluminescent trees and floating lanterns",
width: 1024,
height: 1024,
numInferenceSteps: 50,
guidanceScale: 5.0,
seed: 42);
Constructors
| Constructor | Summary |
|---|---|
HiDreamModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, MMDiTXNoisePredictor<T> predictor, StandardVAE<T> vae, IConditioningModule<T> conditioner, HiDreamVariant variant, int? seed) |
Initializes a new instance of HiDreamModel. |
Properties
| Property | Summary |
|---|---|
HiDreamVariant Variant { get; } |
Gets the model variant. |
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) |
|
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 HiDream (1024x1024). |
int DefaultWidth |
Default image width for HiDream (1024x1024). |