DenoisedScoreDistillation<T>

Models & Types · AiDotNet.Diffusion.Distillation

Denoised Score Distillation (DSD) for artifact-free 3D generation.

For Beginners

Regular SDS compares noise predictions, which can be imprecise. DSD takes a different approach — it fully generates a clean reference image and compares it to what the 3D model renders. This is like having the teacher paint a complete example for every critique, giving much clearer feedback to the 3D model.

How It Works

DSD addresses the mode-seeking bias of SDS by using a fully denoised reference image instead of comparing noisy predictions. At each optimization step, it runs the full diffusion denoising chain to get a clean reference, then uses the difference between this reference and the rendered view as the gradient. While more expensive per step, DSD produces significantly sharper and more diverse 3D results.

Reference: Hertz et al., “Denoised Score Distillation”, 2024

Constructors

Constructor Summary
DenoisedScoreDistillation(IDiffusionModel<T> diffusionModel, double guidanceScale, int denoisingSteps) Initializes a new DSD instance.

Properties

Property Summary
double GuidanceScale { get; } Gets the guidance scale.
int DenoisingSteps { get; } Gets the number of full denoising steps for reference generation.

Methods

Method Summary
Vector<T> ComputeGradient(Vector<T> denoisedReference, Vector<T> renderedView, double weight) Computes the DSD gradient from a fully denoised reference.