Table of Contents

Enum DiffusionPredictionType

Namespace
AiDotNet.Enums
Assembly
AiDotNet.dll

Defines what the diffusion model predicts during the denoising process.

public enum DiffusionPredictionType

Fields

Epsilon = 0

Model predicts the noise (epsilon) that was added to the clean sample.

This is the most common prediction type, used in the original DDPM paper. The model learns: given x_t, predict epsilon such that x_t = sqrt(alpha_cumprod) * x_0 + sqrt(1-alpha_cumprod) * epsilon

For Beginners: The model looks at a noisy image and guesses "what noise was added to make it look like this?" This is like asking someone to identify what static was overlaid on a TV signal.

Sample = 1

Model directly predicts the clean sample (x_0).

Instead of predicting noise, the model directly predicts what the original clean sample looked like. This can sometimes provide more direct gradients but may be harder to train.

For Beginners: The model looks at a noisy image and directly guesses "what was the original clear image?" This is like asking someone to imagine what a static-filled TV picture would look like if it were clear.

VPrediction = 2

Model predicts v = sqrt(alpha_cumprod) * epsilon - sqrt(1-alpha_cumprod) * x_0.

V-prediction is a velocity-based parameterization that can provide more stable training, especially for continuous-time diffusion models. It's used in some advanced diffusion models like Imagen.

For Beginners: This is a mathematical trick that combines noise and sample prediction in a way that makes training more stable. Think of it as predicting "the direction of change" rather than either endpoint.

Remarks

Different diffusion models can be trained to predict different targets. The prediction type affects how the scheduler interprets the model output and computes the denoised sample.

For Beginners: When a model looks at a noisy image, it can be trained to predict:

  • Epsilon (noise): "What noise was added to make this blurry?"
  • Sample: "What did the clean image look like?"
  • V-prediction: A mathematical blend of both (more stable training)

Most models use Epsilon prediction as it's the most common and well-studied approach.

Edit this page