SANAModel<T>
Models & Types · AiDotNet.Diffusion.TextToImage
SANA model for efficient high-resolution text-to-image generation.
Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → SANAModel<T>
For Beginners
SANA generates high-quality images very efficiently. While other models need billions of parameters, SANA achieves similar quality with only 600 million, making it much faster and requiring less memory.
How SANA works:
- Text is encoded by a Gemma 2B language model (Google’s text encoder)
- A 20-layer linear DiT processes text embeddings and noise with O(n) attention
- DC-AE with 32x spatial compression decodes tiny latents to full images
- Flow matching training enables efficient 20-step generation
Model variants:
- SANA-0.6B: Default, 600M parameter DiT
- SANA-1.6B: Larger 1.6B parameter variant for higher quality
Key characteristics:
- 0.6B parameters in the transformer (vs 2-12B for competitors)
- 32x spatial compression via DC-AE (vs 8x standard)
- Linear attention: O(n) instead of O(n^2) quadratic attention
- Gemma 2B text encoder for strong multilingual prompt understanding
- 32 latent channels for high information retention
- Up to 4096x4096 resolution generation
Advantages:
- Extremely efficient: 100x+ faster than Flux1/SDXL for similar quality
- Low VRAM: runs on consumer GPUs with 8GB VRAM
- High resolution: native 4K generation support
- Strong text rendering via Gemma encoder
Limitations:
- Less community ecosystem than Stable Diffusion family
- Fewer LoRA/ControlNet adaptations available
- 32x compression can lose very fine details at low resolution
How It Works
SANA uses a linear DiT architecture with efficient linear attention for high-resolution generation (up to 4K). It achieves state-of-the-art quality with a compact 0.6B parameter model through Deep Compression Autoencoder (DC-AE) with 32x spatial compression and linear attention mechanisms that reduce compute from O(n^2) to O(n).
Technical specifications:
- Architecture: Linear DiT with efficient linear attention
- Transformer: 0.6B params, hidden 2240, 20 layers, 20 heads
- Text encoder: Gemma 2B (2048-dim embeddings)
- VAE: DC-AE with 32 latent channels, 32x spatial compression
- Training: Flow matching with linear noise schedule
- Default: 20 inference steps, guidance scale 4.5
- Resolution: 512x512 to 4096x4096 (aspect-ratio aware)
Reference: Xie et al., “SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers”, 2024
Example
// Create default SANA 0.6B
var sana = new SANAModel<float>();
// Create larger 1.6B variant
var sanaLarge = new SANAModel<float>(variant: SANAVariant.Large);
// Generate an image
var image = sana.GenerateFromText(
prompt: "A serene mountain lake at sunset with reflections",
width: 1024,
height: 1024,
numInferenceSteps: 20,
guidanceScale: 4.5,
seed: 42);
Constructors
| Constructor | Summary |
|---|---|
SANAModel(NeuralNetworkArchitecture<T> architecture, DiffusionModelOptions<T> options, INoiseScheduler<T> scheduler, EMMDiTPredictor<T> predictor, DeepCompressionVAE<T> vae, IConditioningModule<T> conditioner, SANAVariant variant, int? seed) |
Initializes a new instance of SANAModel with full customization support. |
Properties
| Property | Summary |
|---|---|
IConditioningModule<T> Conditioner { get; } |
|
INoisePredictor<T> NoisePredictor { get; } |
|
IVAEModel<T> VAE { get; } |
|
SANAVariant Variant { get; } |
Gets the model variant. |
int CompressionFactor { get; } |
Gets the spatial compression factor of the DC-AE (32x). |
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 SANA (1024x1024). |
int DefaultWidth |
Default image width for SANA (1024x1024). |