HeunDiscreteScheduler<T>

Models & Types · AiDotNet.Diffusion.Schedulers

Heun discrete scheduler for diffusion model sampling using second-order Heun’s method.

Inheritance: NoiseSchedulerBase<T> → HeunDiscreteScheduler<T>

For Beginners

Heun’s method is like Euler but smarter:

  1. Euler takes one step and hopes for the best
  2. Heun takes a trial step, evaluates the derivative there too,

then averages both derivatives for a more accurate step

Key characteristics:

  • Second-order accuracy (better than Euler per step)
  • Two model evaluations per step (so 20 Heun steps = 40 model calls)
  • Good quality with fewer steps than Euler
  • Deterministic: same seed always produces the same result

When to use Heun:

  • When you want higher quality per step than Euler
  • When model evaluation cost is acceptable
  • When you want smooth, accurate trajectories

How It Works

The Heun scheduler implements second-order ODE solving for diffusion sampling. It performs two model evaluations per step (predictor + corrector) to achieve higher accuracy than first-order methods like Euler.

Reference: Karras et al., “Elucidating the Design Space of Diffusion-Based Generative Models”, NeurIPS 2022

Constructors

Constructor Summary
HeunDiscreteScheduler(SchedulerConfig<T> config) Initializes a new instance of the Heun discrete scheduler.

Methods

Method Summary
HeunDiscreteScheduler<T> CreateDefault() Creates a Heun scheduler with default Stable Diffusion settings.
Vector<T> Step(Vector<T> modelOutput, int timestep, Vector<T> sample, T eta, Vector<T> noise) Performs one Heun denoising step (second-order method).
void SetTimesteps(int inferenceSteps) Sets up the inference timesteps and computes sigma schedule.