SDUpscalerModel<T>
Models & Types · AiDotNet.Diffusion.SuperResolution
Stable Diffusion x4 Upscaler model for text-guided latent super-resolution.
Inheritance: DiffusionModelBase<T> → LatentDiffusionModelBase<T> → SDUpscalerModel<T>
For Beginners
This model increases image resolution by 4x with text guidance.
How the SD Upscaler works:
- The low-resolution image is resized and concatenated with latent noise (7 input channels)
- A text prompt guides what details to add during upscaling
- The U-Net denoises in latent space over 75 steps
- The VAE decodes the result to a 4x larger image
Key characteristics:
- 4x upscaling (128 to 512, 256 to 1024, etc.)
- Text-guided: prompts control what details are added
- 7 input channels: 4 latent + 3 low-res RGB conditioning
- Based on SD 2.x architecture with OpenCLIP text encoder
- DDIM scheduler for efficient 75-step inference
When to use SD Upscaler:
- Upscaling images with specific desired detail types
- Combining upscaling with style transfer
- When you want prompt control over the upscaling process
- AI-generated image enhancement
Limitations:
- Fixed 4x upscale factor
- Slower than Real-ESRGAN due to 75 inference steps
- May add details inconsistent with original content at high guidance
How It Works
The Stable Diffusion x4 Upscaler takes a low-resolution image and upscales it 4x using a latent diffusion process conditioned on both the low-resolution input (concatenated in latent space) and a text prompt for guided detail generation.
Architecture components:
- U-Net with 7 input channels (4 latent + 3 low-res RGB conditioning)
- 320 base channels with [1, 2, 4, 4] multipliers (SD 2.x backbone)
- OpenCLIP ViT-H/14 text encoder (1024-dim cross-attention)
- Standard VAE (4-channel latent space, 0.18215 scale factor)
- DDIM scheduler for faster inference
Technical specifications:
- Architecture: U-Net with low-res RGB conditioning
- Input channels: 7 (4 latent noise + 3 low-res RGB)
- Output channels: 4 (latent space)
- Base channels: 320, multipliers [1, 2, 4, 4]
- Cross-attention dimension: 1024 (OpenCLIP ViT-H/14)
- 2 residual blocks per level
- Attention at 4x, 2x, and 1x downsampling
- Noise schedule: Linear beta [0.0001, 0.02], 1000 timesteps
- Default inference steps: 75
- Default guidance scale: 7.5
- Upscale factor: 4x
Reference: Rombach et al., “Stable Diffusion x4 Upscaler”, Stability AI, 2022
Example
var lowResImage = Tensor<float>.CreateRandom(1, 3, 32, 32);
// Create with industry-standard defaults
var upscaler = new SDUpscalerModel<float>();
// Text-guided 4x upscaling
var upscaled = upscaler.ImageToImage(
inputImage: lowResImage,
prompt: "high resolution, sharp details, 4k photograph",
negativePrompt: "blurry, pixelated, noisy",
strength: 0.75,
numInferenceSteps: 75,
guidanceScale: 7.5);
// Simple generation for testing
var result = upscaler.GenerateFromText(
prompt: "a detailed landscape photograph",
width: 512,
height: 512);
Constructors
| Constructor | Summary |
|---|---|
SDUpscalerModel(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 SDUpscalerModel with full customization support. |
Properties
| Property | Summary |
|---|---|
IConditioningModule<T> Conditioner { get; } |
|
INoisePredictor<T> NoisePredictor { get; } |
|
IVAEModel<T> VAE { get; } |
|
int LatentChannels { get; } |
|
int UpscaleFactor { get; } |
Gets the upscale factor (4x). |
Methods
| Method | Summary |
|---|---|
ModelMetadata<T> GetModelMetadata() |
|
Tensor<T> GenerateFromText(string prompt, string negativePrompt, int width, int height, int numInferenceSteps, double? guidanceScale, int? seed) |
Generates an image from a text prompt using SD Upscaler defaults. |
Tensor<T> ImageToImage(Tensor<T> inputImage, string prompt, string negativePrompt, double strength, int numInferenceSteps, double? guidanceScale, int? seed) |
Performs text-guided 4x image upscaling. |
Fields
| Field | Summary |
|---|---|
int DefaultHeight |
Default output image height (512). |
int DefaultWidth |
Default output image width (512). |