VariationalScoreDistillation<T>
Models & Types · AiDotNet.Diffusion.Distillation
Variational Score Distillation (VSD) for high-fidelity text-to-3D generation.
For Beginners
SDS sometimes produces washed-out or over-saturated 3D models. VSD fixes this by training a second model that learns what the 3D model’s renders look like. By comparing the “ideal” score (from the pretrained model) with the “current” score (from the adapted model), it gives more precise, less noisy feedback for 3D optimization.
How It Works
VSD improves upon SDS by reducing the over-saturation and over-smoothing artifacts. It maintains a separate LoRA-adapted diffusion model trained on rendered views and uses the difference between the pretrained model’s score and the LoRA model’s score as the gradient. This effectively performs variational inference in the diffusion model’s data space.
Reference: Wang et al., “ProlificDreamer: High-Fidelity and Diverse Text-to-3D Generation with Variational Score Distillation”, NeurIPS 2023
Constructors
| Constructor | Summary |
|---|---|
VariationalScoreDistillation(IDiffusionModel<T> pretrainedModel, double guidanceScale, double loraLearningRate, int loraRank) |
Initializes a new VSD instance. |
Properties
| Property | Summary |
|---|---|
double GuidanceScale { get; } |
Gets the guidance scale for the pretrained model. |
int LoRARank { get; } |
Gets the LoRA rank for the particle model. |
Methods
| Method | Summary |
|---|---|
T ComputeParticleLoss(Vector<T> particlePrediction, Vector<T> targetNoise) |
Computes the LoRA training loss for the particle model. |
Vector<T> ComputeGradient(Vector<T> pretrainedScore, Vector<T> particleScore, double timestepWeight) |
Computes the VSD gradient using both pretrained and particle model scores. |