IntervalScoreMatching<T>

Models & Types · AiDotNet.Diffusion.Distillation

Interval Score Matching (ISM) for improved 3D score distillation.

For Beginners

SDS gradients are noisy because they compare random noise with the model’s prediction. ISM is cleverer — it uses a deterministic process to add and remove noise over a small interval. This gives much smoother, more reliable feedback for 3D optimization, resulting in cleaner 3D models with fewer artifacts.

How It Works

ISM reduces the variance of SDS gradients by using deterministic DDIM inversion between two timesteps instead of random noise injection. Given a rendered view, ISM deterministically noises it to timestep t2, then denoises to timestep t1, and uses the difference as the gradient signal. This “interval” approach produces much lower-variance gradients than SDS.

Reference: Liang et al., “LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching”, CVPR 2024

Constructors

Constructor Summary
IntervalScoreMatching(IDiffusionModel<T> diffusionModel, double guidanceScale, int intervalSteps) Initializes a new ISM instance.

Properties

Property Summary
double GuidanceScale { get; } Gets the guidance scale.
int IntervalSteps { get; } Gets the number of DDIM steps in the interval.

Methods

Method Summary
Vector<T> ComputeGradient(Vector<T> denoisedAtT1, Vector<T> invertedAtT2, Vector<T> originalRender, double timestepWeight) Computes the ISM gradient between two timestep points.