Table of Contents

Class DDIMScheduler<T>

Namespace
AiDotNet.NeuralNetworks.Diffusion.Schedulers
Assembly
AiDotNet.dll

DDIM (Denoising Diffusion Implicit Models) scheduler implementation.

public sealed class DDIMScheduler<T> : NoiseSchedulerBase<T>, INoiseScheduler<T>

Type Parameters

T

The numeric type used for calculations.

Inheritance
DDIMScheduler<T>
Implements
Inherited Members

Remarks

DDIM is a faster variant of DDPM that can achieve similar quality with far fewer denoising steps. While DDPM requires many steps (often 1000), DDIM can achieve similar quality with 50 or fewer steps by using a different mathematical formulation.

For Beginners: DDIM is like a shortcut for removing noise from images.

Imagine you have a very blurry photo and need to make it clear:

  • DDPM (original method): Take 1000 tiny steps to slowly reveal the image
  • DDIM (this method): Take 50 larger steps to reveal the image faster

The magic is the "eta" parameter:

  • eta=0: Deterministic - same input always produces same output (faster, consistent)
  • eta=1: Stochastic - adds randomness like DDPM (slower, more variety)
  • eta between 0-1: Mix of both behaviors

Key advantages of DDIM:

  • Much faster generation (10-50x fewer steps needed)
  • Deterministic option allows reproducible results
  • Can interpolate smoothly between images (useful for animations)

Reference: "Denoising Diffusion Implicit Models" by Song et al., 2020

Constructors

DDIMScheduler(SchedulerConfig<T>)

Initializes a new instance of the DDIM scheduler.

public DDIMScheduler(SchedulerConfig<T> config)

Parameters

config SchedulerConfig<T>

Configuration for the scheduler including beta schedule parameters.

Remarks

For Beginners: Create a DDIM scheduler with custom settings, or use CreateDefault() for standard DDPM defaults. 

// Create with default settings
var config = SchedulerConfig<double>.CreateDefault();
var scheduler = new DDIMScheduler<double>(config);

// Set up for 50 inference steps
scheduler.SetTimesteps(50);

Exceptions

ArgumentNullException

Thrown when config is null.

Methods

Step(Vector<T>, int, Vector<T>, T, Vector<T>?)

Performs one DDIM denoising step.

public override Vector<T> Step(Vector<T> modelOutput, int timestep, Vector<T> sample, T eta, Vector<T>? noise = null)

Parameters

modelOutput Vector<T>

The model's noise prediction (epsilon).

timestep int

The current timestep in the diffusion process.

sample Vector<T>

The current noisy sample.

eta T

Stochasticity parameter:

  • 0 = fully deterministic (pure DDIM)
  • 1 = fully stochastic (equivalent to DDPM)
  • Values in between interpolate the behavior
noise Vector<T>

Optional noise for stochastic sampling. Required when eta > 0 for true stochastic behavior. If null and eta > 0, falls back to deterministic (zero noise).

Returns

Vector<T>

The denoised sample for the previous timestep.

Remarks

The DDIM step formula (simplified): 1. Predict original sample: x_0 = (x_t - sqrt(1-alpha_cumprod) * eps) / sqrt(alpha_cumprod) 2. Compute "direction pointing to x_t": d = sqrt(1-alpha_prev - sigma^2) * eps 3. Compute previous sample: x_{t-1} = sqrt(alpha_prev) * x_0 + d + sigma * noise

For Beginners: This function takes your current noisy image and the model's guess of what noise is in it, then removes some of that noise to get a cleaner image. The eta parameter controls whether this removal is exact (eta=0) or has some randomness (eta>0).

Exceptions

ArgumentNullException

Thrown when modelOutput or sample is null.

ArgumentException

Thrown when modelOutput and sample have different lengths.

ArgumentOutOfRangeException

Thrown when timestep is out of range.

Edit this page