RewardScoreDistillation<T>

Models & Types · AiDotNet.Diffusion.Distillation

Reward-weighted Score Distillation Sampling (RewardSDS) for preference-aligned 3D generation.

For Beginners

Standard SDS creates realistic-looking 3D objects, but they may not match what people actually want aesthetically. RewardSDS adds a “quality judge” (reward model) that steers the generation toward more visually appealing results. It’s like having both a realism expert and an art critic guiding the 3D artist simultaneously.

How It Works

RewardSDS augments standard SDS gradients with reward model feedback. Instead of using only the diffusion model’s score, it weights the gradient by a reward signal that measures alignment with human preferences (aesthetics, text-image correspondence, etc.). This produces 3D objects that not only look realistic but also match desired aesthetic qualities.

Reference: Adapted from DreamReward (Ye et al., 2024) combining human preference alignment with score distillation for text-to-3D generation

Constructors

Constructor Summary
RewardScoreDistillation(IDiffusionModel<T> diffusionModel, double guidanceScale, double rewardWeight) Initializes a new RewardSDS instance.

Properties

Property Summary
double GuidanceScale { get; } Gets the guidance scale for score distillation.
double RewardWeight { get; } Gets the reward signal weight.

Methods

Method Summary
Vector<T> ComputeGradient(Vector<T> predictedNoise, Vector<T> addedNoise, Vector<T> rewardGradient, double timestepWeight) Computes the reward-weighted SDS gradient.