RefinerStage<T>
Models & Types · AiDotNet.Diffusion.Alignment
Refiner stage for late-stage noise-add-then-denoise detail improvement.
For Beginners
Imagine you’ve painted a picture but want to add more detail. The refiner slightly blurs the image (adds noise), then carefully sharpens it back. This re-processing enhances fine details like skin texture, fabric patterns, and small objects that the main generation pass might have missed.
How It Works
The refiner stage takes a nearly-finished generated image, adds a small amount of noise back, and re-denoises using a specialized refiner model (or the same model with different conditioning). This late-stage refinement enhances fine details, textures, and overall coherence. Used in SDXL’s two-stage pipeline and other multi-stage generation approaches.
Reference: Podell et al., “SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis”, ICLR 2024 (refiner stage)
Constructors
| Constructor | Summary |
|---|---|
RefinerStage(IDiffusionModel<T> refinerModel, double strength, double guidanceScale, int refinerSteps) |
Initializes a new refiner stage. |
Properties
| Property | Summary |
|---|---|
double GuidanceScale { get; } |
Gets the classifier-free guidance scale for the refiner. |
double Strength { get; } |
Gets the noise strength for refinement (0.0 = no change, 1.0 = full re-generation). |
int RefinerSteps { get; } |
Gets the number of denoising steps the refiner uses. |
Methods
| Method | Summary |
|---|---|
T ComputeRefinementDelta(Vector<T> beforeRefinement, Vector<T> afterRefinement) |
Computes the refinement quality improvement estimate. |
Vector<T> AddNoiseForRefinement(Vector<T> cleanLatent, Vector<T> noise) |
Adds noise to a latent at the specified strength level. |
int GetStartTimestep(int totalTimesteps) |
Computes the starting timestep for the refiner based on strength. |