ScoreDistillationSampling<T>

Models & Types · AiDotNet.Diffusion.Distillation

Score Distillation Sampling (SDS) for text-to-3D and generator optimization.

For Beginners

SDS uses a 2D image generator as a “critic” for 3D models. It renders the 3D model from a random angle, asks the 2D model “how could this image be improved?”, and uses that feedback to update the 3D model. Repeating this from many angles creates a 3D object that looks good from every direction.

How It Works

SDS computes gradients from a pretrained 2D diffusion model to optimize a 3D representation (NeRF, mesh, etc.). It adds noise to rendered views, asks the diffusion model to predict the noise, and uses the difference between added and predicted noise as a gradient signal. This enables text-to-3D generation without any 3D training data.

Reference: Poole et al., “DreamFusion: Text-to-3D using 2D Diffusion”, ICLR 2023

Constructors

Constructor Summary
ScoreDistillationSampling(IDiffusionModel<T> diffusionModel, double guidanceScale, double minTimestep, double maxTimestep, double gradientScale) Initializes a new SDS instance.

Properties

Property Summary
double GuidanceScale { get; } Gets the classifier-free guidance scale.

Methods

Method Summary
Vector<T> ComputeGradient(Vector<T> predictedNoise, Vector<T> addedNoise, double timestepWeight) Computes the SDS gradient for a rendered view.
int SampleTimestep(int numTimesteps, Random random) Samples a random timestep within the configured range.